okf-mcp
okf-mcp publishes Open Knowledge Format v0.2 from the systems that already hold the knowledge, and serves the result to agents. It is a local-first consumer, validator, graph index, CLI, and MCP server.
The part that is hard is writing to a catalog safely. A configured producer reads an external source and returns a candidate bundle; okf-mcp validates it, then publishes it under an ownership manifest that records exactly which files it owns. Hand-authored content is never adopted, a source that did not change rewrites nothing, and no producer output escapes its bundle. @mfdaves/okf-postgres reads PostgreSQL catalogs this way without reading a single table row.
Reading is the ordinary half. It consumes an OKF bundle directory of Markdown files with YAML frontmatter. An optional workspace mode can federate several bundles. Concepts are exposed through CLI commands and MCP resources and tools for validation, structured search, graph navigation, provenance inspection, and proposal-based authoring.
The core intentionally has no database, embeddings, build step, or hosted-service dependency. It uses js-yaml for safe YAML, CommonMark for Markdown structure, MiniSearch for in-memory BM25+ text retrieval, and the official Model Context Protocol TypeScript SDK v2 for MCP over stdio and Streamable HTTP. Local root mode makes no network calls. Optional remote loading fetches public Markdown concepts and only their explicitly referenced inert assets from GitHub. Nothing in the v0.2 computation support executes code or attests a receipt.
OKF v0.2 Support And Extensions
OKF v0.2 intentionally specifies a portable file format, not a serving or query runtime. okf-mcp keeps that boundary explicit:
| Area | Official OKF v0.2 | okf-mcp behavior |
|---|
| Bundle and identity | A directory tree of Markdown files; a Concept ID is its bundle-relative path without .md | --root maps directly to one bundle; okf:// is an optional workspace locator, not the portable Concept ID |
| Concept metadata | Required type; recommended title, description, resource, and tags; unknown keys are allowed | Preserves extension fields and unknown types while reporting normative conformance separately from workspace policy |
| Provenance and lifecycle | sources, usage_window, generated, verified, status, and stale_after | Normalizes these fields for search, provenance traversal, trust tiers, and deterministic freshness checks |
| References | Markdown links and path-valued resource, sources[].resource, computation, executor.resource, and attester.resource fields | Builds graph edges and bounded inert asset snapshots without executing or implicitly fetching referenced code |
| Attested Computation | Defines contract fields and an informative consumer flow while deferring runtime wire protocols and attester packaging | Statically inspects contracts and digests, checks declared parameter and receipt field names, and never executes or claims attestation |
| v0.1 compatibility | Allows timestamp fallback when the whole generated mapping is absent and # Citations fallback when the sources key is absent | Consumes both forms and adds review-only migration checks and proposals |
The following are okf-mcp extensions rather than requirements of the format:
- CLI, MCP, and HTTP interfaces; in-memory search and graph views
- optional multi-bundle
okf.project.yaml workspaces and typed relations
- compatibility
id, aliases, and okf:// locators
- proposal-backed authoring with explicit acceptance
- bounded GitHub remote loading and explicitly mapped pinned Git sources
- generator plugins and stricter opt-in project policies such as
strictLinks
Install And Run
Node 22 or newer is required.
Install from the GitHub release:
git clone --branch v0.12.0-rc.1 https://github.com/mfdaves/okf-mcp.git
cd okf-mcp
npm ci
node bin/okf-mcp.js --root ./path/to/okf validate
Pin the published version for reproducible use:
npx -y @mfdaves/okf-mcp@0.12.0-rc.1 --version
npx -y @mfdaves/okf-mcp@0.12.0-rc.1 --root ./path/to/okf validate
For a persistent installation:
npm install --global @mfdaves/okf-mcp@0.12.0-rc.1
okf --version
okf --root ./path/to/okf validate
okf-mcp --root ./path/to/okf mcp
To work from the current source branch:
git clone https://github.com/mfdaves/okf-mcp.git
cd okf-mcp
npm ci
npm test
node bin/okf-mcp.js --version
--root accepts one local OKF bundle directory and is the recommended okf-mcp interface for a single bundle. The portable identity of each concept is its extensionless path inside that root.
--bundle accepts either a path or id=path. Multiple flags remain supported for compatibility. --project and its bundles: list are an optional okf-mcp federation/authoring extension, not part of OKF v0.2.
--remote-bundle accepts id=https://github.com/<owner>/<repo>/tree/<ref>/<path>. It fetches public Markdown first, then only bundle-local files explicitly named by standard v0.2 resource fields. Remote content remains read-only and inert.
--inspect prints a compact graph summary and exits. Without --inspect and without an explicit command, the process starts a stdio MCP server.
The package exposes both okf and okf-mcp binaries when installed. Without an explicit source, the CLI first discovers the nearest root index.md declaring okf_version; nearest-project discovery remains a compatibility fallback.
CLI exit statuses are 0 for success, 1 for validation or operational failure, and 2 for invalid usage. Unknown options are rejected.
Included OKF Reference
This repository publishes a self-describing OKF bundle for the product, its runtime boundaries, interfaces, authoring workflows, and safety policy. Its portable entry Concept ID is overview/okf-mcp; okf://okf-mcp/overview/okf-mcp remains the workspace/MCP resource locator.
Validate and query the bundled reference from a checkout or installed package:
okf --root okf/okf-mcp validate
okf --root okf/okf-mcp search "proposal"
okf --root okf/okf-mcp concept overview/okf-mcp
Load the reference bundle directly from the current source branch:
okf --remote-bundle okf-mcp=https://github.com/mfdaves/okf-mcp/tree/main/okf/okf-mcp --inspect
The published v0.8.0 tag retains its original
okf/bundles/okf-mcp path. The shorter okf/okf-mcp path was prepared in the
unpublished 0.8.1 maintenance checkpoint and ships with v0.9.0; the
immutable v0.8.0 release remains unchanged.
The @mfdaves/okf-mcp npm package includes both okf.project.yaml and the
complete reference bundle.
Optional Multi-Bundle Project Config
Use okf.project.yaml only when one process must federate multiple roots, configure generators or external metadata producers, or enforce a project-wide relation vocabulary:
project: Example
strictLinks: false
bundles:
- id: app
root: okf/bundles/app
include: ["**/*.md"]
exclude: ["archive/**"]
- id: data
root: okf/bundles/data
- id: company-database
root: okf/company-database
relationTypes:
- deployed_by
remoteBundles:
- id: shared
url: https://github.com/example/okf-atlas/tree/main/bundles/shared
include: ["public/**"]
exclude: ["drafts/**"]
plugins:
- name: docs
type: filesystem
root: docs
output: okf/bundles/app/generated/docs
bundle: app
producers:
- name: main-database
type: postgresql
package: "@mfdaves/okf-postgres"
bundle: company-database
config:
connectionEnv: DATABASE_URL
source: company-db
schemas: [public]
includeIndexes: true
Run project commands:
okf --project okf.project.yaml validate
okf --project okf.project.yaml search "orders"
okf --project okf.project.yaml graph mermaid
okf --project okf.project.yaml generate
okf --project okf.project.yaml producer list
okf --project okf.project.yaml producer preview main-database
okf --project okf.project.yaml producer run main-database --write --actor process:nightly-refresh
okf --project okf.project.yaml mcp
okf --project okf.project.yaml mcp --authoring
okf --project okf.project.yaml mcp --allow-remote-tool
OKF_WRITE_TOKEN=change-me okf --project okf.project.yaml serve
okf --remote-bundle shared=https://github.com/example/okf-atlas/tree/main/bundles/shared --inspect
Commands:
mcp
validate
graph [json|dot|mermaid]
search <query>
concept <concept-id-or-locator>
neighbors <concept-id-or-locator>
paths <from> <to>
provenance <uri>
edge-kinds
computation inspect|prepare|check-receipt
asset <okf-asset-uri>
source <concept-id-or-locator> <source-id>
migrate check|preview
generate
producer list|preview|run
serve
Producers are separate from the built-in plugins: generators. A producer reads metadata from an external source and returns an in-memory candidate OKF bundle; okf-mcp validates and publishes that candidate. Install each producer package alongside the project, then allowlist its exact bare package name in the trusted project configuration. MCP callers can select a configured producer but cannot replace its package, configuration, credentials, queries, bundle, or output paths.
The package must be installed under a node_modules directory belonging to the project, resolved by its exact bare name. Ordinary registry installs, npm workspaces, npm link, and store-backed layouts such as pnpm all satisfy this; a path in configuration never does.
The package must export the versioned okfProducer API. Producer API v1 targets OKF v0.2 and separates validateConfig() from side-effect-free destination generation. Candidate concepts require native generated and non-empty sources provenance. The host rejects unsafe paths, malformed concepts, unresolved links or typed relations, unowned file collisions, locally modified managed files, secret values in output, and any candidate that would leave the complete project invalid.
Both interfaces run the same host. Use the CLI for scheduled refreshes:
okf --project okf.project.yaml producer preview main-database
okf --project okf.project.yaml producer run main-database --write --actor process:nightly-refresh
producer preview exits 1 when the candidate is invalid; producer run exits 1 when nothing was published. Both print the same receipt the MCP tools return. A producer that cannot generate may throw an error with its own code (lowercase, ^[a-z][a-z0-9_]{0,63}$, for example access_denied) and a details object of scalar values; the receipt then carries one producer-layer diagnostic with that code and the bounded details (identifier keys, finite numbers, booleans, null, single-line strings up to 200 characters that hold no secret the producer read). The error's message is never reported. An error without such a code is reported as producer_execution_failed.
For a configured project, MCP exposes:
okf_list_producers: list configuration without loading package code or reading the source
okf_preview_producer: read source metadata and return a validated, non-writing diff
okf_run_producer: regenerate under the shared writer queue and publish with revision checks and rollback; available only with --write plus a truthful --actor using human:<id>, process:<id>, or provider/model syntax
Preview and run accept only the configured producer name plus optional receipt detail. Receipts include bounded numeric source-object summaries and publication-diff counts. Publication records owned relative paths and SHA-256 digests in the bundle-local .okf-producer.json. Only unchanged paths owned by the prior manifest can be updated or removed; hand-authored content is never adopted implicitly. The manifest contains no producer configuration, credentials, or connection details.
Two rules keep a managed bundle stable across repeated runs:
- A producer restamps
generated.at on every run. The host compares each candidate against the published file ignoring that stamp, so a source that did not change reports unchanged and leaves the bytes, the timestamps, and the manifest untouched. Publication-diff counts therefore describe the source, not the run.
- A manifest entry alone is not authority to destroy work. Any file the host would rewrite or remove must itself carry
generated provenance, so a damaged, hand-merged, or forged manifest cannot make a producer delete hand-authored content. A reserved index.md or log.md cannot carry generated frontmatter, so for those the proof is that their bytes still match the manifest's digest; a hand-edited one stops the run with producer_owned_file_modified. A refusal names the file in its diagnostic's details.path.
Directories emptied by a stale deletion are removed with their last managed file.
@mfdaves/okf-postgres is the PostgreSQL implementation. Its average-use v1 covers database, schema, table/view, column, key, check, enum, index, and description metadata without reading table rows. Keep connection strings in an environment variable named by connectionEnv; do not place a connection string in YAML.
serve options:
--host <host>: bind host, default 127.0.0.1
--port <port>: bind port, default 8765
--write-token <token>: bearer token for write endpoints; defaults to OKF_WRITE_TOKEN
--proposal-root <path>: proposal JSON directory; defaults to .okf-proposals under the selected local root or project
Proposals are stored state, not a derived cache, so the proposal directory is never cleaned up automatically. A server started with --git-commit refuses to publish while the worktree has untracked files, so that directory must be ignored by Git: add .okf-proposals/ to .gitignore (as this repository does), or pass --proposal-root with a directory Git already ignores. In project mode the proposal root must stay inside the project root.
MCP Client Config
The npm-based examples below use the current published release. A source checkout can invoke its executable bin/okf-mcp.js with the same arguments.
Example client configuration:
{
"mcpServers": {
"okf": {
"command": "npx",
"args": [
"-y",
"@mfdaves/okf-mcp@0.12.0-rc.1",
"--root",
"/absolute/path/to/okf",
"mcp"
]
}
}
}
Project config mode, with read-only project helpers but without proposal mutations:
{
"mcpServers": {
"okf": {
"command": "npx",
"args": [
"-y",
"@mfdaves/okf-mcp@0.12.0-rc.1",
"--project",
"/absolute/path/to/repo/okf.project.yaml",
"mcp"
]
}
}
}
Add --authoring to enable proposal creation, acceptance, and rejection. For a smaller direct-write surface, add --write --actor <actor> to expose read-only okf_validate_changes and validated-batch okf_apply_changes; add --git-commit to commit each successful batch when the catalog is in a clean Git worktree. Add --allow-remote-tool to let MCP clients load arbitrary supported public remote bundles at runtime. Configured remote bundles remain readable without that runtime-loading flag.
The stdio server uses @modelcontextprotocol/server v2. It serves the modern
2026-07-28 MCP revision and the SDK's compatibility path for 2025-era
clients, including 2025-11-25. The SDK owns protocol negotiation, framing,
resource dispatch, tool dispatch, and advertised-schema validation.
Expected failures from a known tool, such as a missing concept, a read-only
bundle, a failed remote fetch, invalid arguments, or a proposal conflict, are
returned as MCP tool results with isError: true. Calls to tools that are not
enabled are rejected by SDK dispatch. Unexpected implementation errors are
masked instead of exposing internal details.
MCP over HTTP
From 0.13.0-rc.1 the mcp command can serve the same tools, resources and
schemas over Streamable HTTP instead of stdio:
okf --root /path/to/okf mcp --http
| Flag | Meaning |
|---|
--http | Serve MCP over Streamable HTTP. |
--host <host> | Address to bind (default 127.0.0.1). A non-loopback address requires --auth. |
--port <port> | Port (default 8766; 0 picks a free one). |
--public-url <url> | The URL clients use for the MCP endpoint. Its path is the endpoint path (default /mcp; served with and without one trailing slash), its host is accepted in Host headers, and with introspection it is the token audience and the advertised resource, as written. |
--allowed-host <host> | Another accepted Host header (hostname; the port is ignored); repeatable. localhost, 127.0.0.1 and [::1] are always accepted. |
--allowed-origin <origin> | Another accepted browser Origin (hostname or origin URL); repeatable. Requests without an Origin header are not affected. |
--auth tokens:<file>, --auth introspect:<url> | Bearer authentication, see below. |
--authorization-server <issuer> | With introspection, the issuer named in the protected-resource metadata. |
--refresh-interval <ms> | Check the files for changes made outside okf-mcp before a tool call or resource read at most this often (default 2000; 0 checks before every one). |
The server answers:
- the MCP endpoint, through the SDK's
createMcpHandler: protocol revision
2026-07-28, and 2025-era clients (including the 2025-11-25 initialize
handshake) through its stateless fallback. No session is kept: GET and
DELETE on the endpoint answer 405 to 2025-era clients.
GET /health with {"status":"ok"}, also beside the endpoint when its
path is deeper (/okf/mcp → /okf/health).
- with introspection, the RFC 9728 protected-resource metadata at
/.well-known/oauth-protected-resource plus the path of --public-url
(for example /.well-known/oauth-protected-resource/mcp) and at
/.well-known/oauth-protected-resource.
Every request is checked against the accepted Host names (403
otherwise); requests to the MCP endpoint are also checked against the accepted
origins (403, without CORS headers) and refused above 1 MiB (413). For an
accepted browser origin the endpoint answers CORS: an OPTIONS preflight gets
204 without authentication, with Access-Control-Allow-Origin (the exact
origin, Vary: Origin), Access-Control-Allow-Methods: GET, POST, DELETE, OPTIONS and Access-Control-Allow-Headers: Authorization, Content-Type, Mcp-Session-Id, MCP-Protocol-Version, Last-Event-ID; every other answer,
401 and 403 included, carries the same Access-Control-Allow-Origin and
Access-Control-Expose-Headers: WWW-Authenticate, Mcp-Session-Id. The
protected-resource metadata is answered the same way (its preflight allows
GET, HEAD, OPTIONS), so a browser client can discover the authorization
server. Each request builds its MCP
server over the one shared catalog state, and tool calls from all clients run
one at a time through one queue.
Principals and scopes
Every HTTP request runs for a principal: a subject, an actor, scopes, and
optionally a name and an email. okf.read is required for any call.
okf.write adds the live write tools (okf_validate_changes,
okf_apply_changes, when the server runs with --write), producer runs
(okf_run_producer) and, over HTTP, producer listing and previews
(okf_list_producers, okf_preview_producer, which run the producer with
the server's own source credentials), the stored proposals (okf_list_proposals,
okf_get_proposal, which can hold complete candidate Markdown) and, with
--allow-remote-tool, load_remote_bundle, which changes the index every
principal reads. Over stdio those flags alone enable them, as before. A principal without okf.write does not see those
tools in tools/list, and calling one fails as calling a disabled tool does.
Writes made for a principal stamp its actor in generated.by and are judged
by the write policy for that actor. With --git-commit, the principal's name
and email become the commit author; a missing one falls back to the
repository's configured Git identity, which also stays the committer. Author
data never fails or blocks a write: <, >, control characters and
surrounding whitespace are removed, a name made only of characters git
refuses (such as ' or ") and an email that is not a plausible address are
dropped, and the dropped field falls back to the configured identity.
Producer runs keep
process:<producer>. --write still requires --actor, which names the
server process as over stdio; every authenticated HTTP principal writes as
its own actor.
Without --auth, every caller is an anonymous read-only principal. That mode
is allowed only on a loopback address, and --write is a usage error there.
--authoring is not available with --http in this release: which scope may
create and accept proposals over HTTP is not decided yet. Over stdio the
principal comes from the flags (--actor, and okf.write only with
--write), so stdio behaves as before.
Token file
--auth tokens:<file> reads a YAML list of tokens by their SHA-256 hash;
plaintext tokens are never stored:
tokens:
- sha256: 3f0c…
subject: ci
actor: process:ci-nightly
scopes: [okf.read, okf.write]
printf %s "$TOKEN" | sha256sum prints the hash. The file must be a regular
file of at most 1 MiB; scopes are okf.read, okf.write or well-formed
names outside okf.* (for tool packs), and unknown fields, other scopes, invalid actors and
repeated hashes are startup errors that name the entry and field, never a
value. Listed tokens do not expire; remove an entry and restart the server to
revoke it.
Token introspection
--auth introspect:<url> checks opaque bearer tokens against an external
OAuth 2.0 authorization server with RFC 7662 introspection. okf-mcp is only a
resource server: it never issues tokens, and clients obtain them from the
authorization server named in the protected-resource metadata.
OKF_INTROSPECTION_CLIENT_ID=okf-mcp OKF_INTROSPECTION_CLIENT_SECRET=… \
okf --root /srv/okf --write --actor process:okf-mcp --git-commit mcp --http \
--public-url https://mcp.example.com/mcp \
--auth introspect:https://auth.example.com/oauth/introspect \
--authorization-server https://auth.example.com
- okf-mcp posts
token=<access token> as a form to the URL, with HTTP Basic
client credentials (form-urlencoded before base64, RFC 6749 section 2.3.1)
read from OKF_INTROSPECTION_CLIENT_ID and
OKF_INTROSPECTION_CLIENT_SECRET only. The URL must be https, or http
on a loopback host, without credentials, query or fragment; redirects are
not followed.
- A token is accepted when
active is true, aud equals --public-url
(one trailing slash ignored on either side), exp is in the future and
scope holds okf.read. The principal's subject
is sub, its actor human:<username> (every character outside
A-Z, a-z, 0-9 and ._@+- replaced by -, so a username can never
take the provider/model form), its scopes every well-formed entry of
scope (okf.read and okf.write gate the core tools, scopes outside
okf.* such as odoo.read gate tool-pack
tools, other okf.* names are ignored), and its git author name and
email, each optional.
- A request without credentials gets
401 with
WWW-Authenticate: Bearer realm="okf-mcp", scope="…", resource_metadata="…"
and no error attribute (RFC 6750 section 3.1); a presented Bearer token
that is refused gets the same challenge with error="invalid_token". The
scope lists every scope the server offers, the same list as
scopes_supported, so the authorization server narrows it to the user's
rights. A token without okf.read gets 403 with
error="insufficient_scope", scope="okf.read". When the authorization server cannot be reached,
times out (5 s) or answers anything but 200 with JSON, the request gets
503 with {"error":"temporarily_unavailable"} and no further detail.
- Answers are cached by the token's SHA-256 in two bounded caches: accepted
ones until the earlier of
exp and 60 seconds (a revoked token keeps
working for up to a minute), refused ones for 30 seconds. At most 8
introspection requests run at once; a request that would need a ninth gets
503 with {"error":"temporarily_unavailable"}.
Token values never appear in a response, an error or a log line, and the
caches keep only their hash; the verified token is handed on only to
tool-pack handlers, as ctx.subjectToken.
Index freshness
Over HTTP the server re-reads files changed outside okf-mcp (for example by
git pull or an editor) before a tool call or a resource read, at most once
per --refresh-interval; initialize, ping and tools/list never trigger
it. The check runs in the call queue, in the same order as the calls. Changes are detected by file size and modification time,
as for writes. An unchanged check costs about 35 ms at 5,000 concepts.
Behind a reverse proxy
- Bind okf-mcp to a loopback address and let the proxy terminate TLS.
- Set
--public-url to the external URL of the endpoint and forward that
path unchanged; forward the health path and
/.well-known/oauth-protected-resource… as well.
- Forward the
Host header as the public host (accepted automatically), or
accept the host the proxy sends with --allowed-host.
- Forward the
Authorization header unchanged.
- Responses can be Server-Sent Events streams: turn response buffering off
for the endpoint and allow long-lived responses.
- One process serves one root or one project; run one process per catalog.
Experimental from 0.13.0-rc.1 and outside the 1.0 freeze: the interface may
change in any release.
A tool pack is an npm package that adds read tools to the MCP server, next to
the core tools. Like a producer, it is listed by exact package name in the
trusted project config and loaded only from the project's node_modules:
toolPacks:
- package: "@example/okf-data-pack"
timeoutMs: 30000
config:
url: https://data.example.com
clientIdEnv: DATA_CLIENT_ID
clientSecretEnv: DATA_CLIENT_SECRET
The package exports:
exports.okfToolPack = {
apiVersion: "1",
id: "data",
validateConfig(config) { return { valid: true, config }; },
serialize: false,
tools: {
data_search: {
description: "…",
annotations: { readOnlyHint: true },
inputSchema: { type: "object", properties: { … }, additionalProperties: false },
outputSchema: { type: "object", … },
scope: "data.read",
async handler(args, ctx) { return { records: [] }; },
},
},
};
- Loading. Packs load when the server starts with
--project, in config
order. A package that is not installed under the project's node_modules,
a descriptor the host does not accept, a configuration validateConfig
refuses, an inline secret in the config (password, secret, token,
apiKey, credential(s)), or a tool name that repeats a core tool or
another pack's tool stops the server with an error naming the package and
tool. Descriptions and schemas come from the package, never from config.
- Listing and access. Pack tools are listed after the core tools, packs
in config order and each pack's tools in the order it declares them. A
principal sees a pack tool when it holds
okf.read and the tool's scope;
calling a tool it does not see fails as calling a disabled tool does. Over
stdio the principal is the process's operator and holds every scope outside
okf.* that the configured packs declare (an okf.write scope still needs
--write). A tool that sets requiresSubjectToken: true (because it acts
for the caller elsewhere) is listed and callable only when the request
carries a subject token, that is over HTTP with --auth; over stdio and
anonymous HTTP it is hidden and refused like a disabled tool.
- Calls. Arguments are validated against
inputSchema exactly like core
tool arguments. ctx holds principal (a frozen copy of subject, actor,
scopes, name and email), subjectToken (over HTTP the caller's verified
bearer token, absent over stdio), config (the normalised config
validateConfig returned, else the configured one; a fresh copy per call),
getSecret(name) (only for environment variables named by config keys
ending in Env) and signal, which aborts when timeoutMs passes or the
request is cancelled. A handler returns a JSON value (at most 1 MiB;
returned as structured content when the tool declares an outputSchema).
- Errors. A handler may throw an error with a snake_case
code and a
details object; the caller receives
{ "error": "Tool execution failed", "tool", "message", "code", "details" }
with a fixed message, the code, and only bounded scalar details: at most
16 entries under identifier keys, values null, booleans, finite numbers or
single-line strings, a string longer than 500 characters cut to 500 ending
in …, and no string holding the caller's token or a secret the pack read.
The text content repeats this for clients that read only text: the fixed
message, then one JSON line with code and details. The error's own
message is never forwarded. Any other failure is
tool_pack_execution_failed; the host's own codes are tool_pack_timeout,
tool_pack_cancelled, tool_pack_result_too_large,
tool_pack_secret_not_allowed and tool_pack_secret_unavailable.
- Concurrency. Pack tools are read tools and run outside the call queue
that serialises knowledge reads and writes, so a slow pack never blocks a
write. A pack that needs the queue sets
serialize: true.
- Scopes over HTTP. The server advertises
okf.read, okf.write and
every scope its packs declare, each once, in config and declaration order:
in the protected-resource metadata's scopes_supported and in the scope
of the 401 challenge, so a client signing in asks for all of them. An
authorization server puts into a token only the scopes asked for at sign-in,
so users who signed in before a pack was added must sign in again to obtain
its scope.
server.json describes the npm package io.github.mfdaves/okf-mcp twice:
as a stdio server, and as a Streamable HTTP server at
http://127.0.0.1:8766/mcp started with the extra --http argument.
Registry-aware clients should prompt for an absolute OKF root path, pass it
through --root, and append the fixed mcp command.
Concept Identity And Extensions
Concept IDs are their bundle-relative Markdown paths with .md removed. This extensionless path is the portable OKF identity. okf-mcp also exposes a workspace-scoped compatibility locator:
okf://<bundle-id>/<extensionless-concept-id>
The former .md URI and a valid custom id remain compatibility lookup aliases. A bare Concept ID resolves only when unique across loaded bundles. For standalone aggregate catalogs, a URI-shaped portable path such as okf://services/queue.md also resolves when services/queue is globally unique and services is not a loaded bundle id. Exact canonical URIs always win; a known-bundle miss or ambiguous portable path stays unresolved. Reserved index.md and log.md resources retain their filenames because they are not concepts.
The id, aliases, and typed relations fields below are okf-mcp extensions. The standard v0.2 identity remains path-derived:
---
id: okf://app/routes/order-status
type: API Route
title: Order Status Route
description: Serves order status state.
aliases: [order-status]
tags: [api, orders]
relations:
- type: consumes
target: okf://data/tables/order_status
- type: configured_by
target: repo://src/routes/order-status.js
---
# Order Status Route
New content should use normal relative or bundle-root Markdown paths for internal links and extension relation targets. Existing okf:// targets remain supported; non-OKF schemes such as repo:// remain opaque compatibility references.
Pinned Git Sources
Code knowledge may live outside the code repository without recording a machine-specific path. Point a standard sources entry at a Git Repository concept and add the okf-mcp git extension below:
sources:
- id: implementation
resource: /repositories/application.md
git:
revision: 0123456789abcdef0123456789abcdef01234567
path: src/application.js
lines: { from: 10, to: 30 }
Map the repository concept only in the local process configuration:
okf --root /path/to/catalog \
--repo repositories/application=/work/application \
source architecture/application implementation
read_git_source and the source CLI command read the pinned blob from the mapped Git object database. They never read the dirty worktree or fetch. Missing mappings and unpinned revisions remain visible but unavailable. Repository mappings may point to normal checkouts, bare repositories, or mounted paths; credentials and local paths stay outside the OKF bundle.
list_bundles
list_concepts
get_concept
search_concepts
list_types
list_tags
list_relation_types
list_edge_kinds
get_provenance
inspect_attested_computation
read_bundle_asset
read_git_source
prepare_attested_computation
check_computation_receipt
check_v02_migration
load_remote_bundle
list_remote_bundles
okf_validate_concept
okf_suggest_concept_path
okf_propose_concept
okf_propose_update
okf_propose_attested_computation
okf_propose_v02_migration
okf_list_proposals
okf_get_proposal
okf_accept_proposal
okf_reject_proposal
okf_validate_changes
okf_apply_changes
okf_list_producers
okf_preview_producer
okf_run_producer
get_graph
get_neighbors
get_subgraph
find_paths
graph_summary
validate_bundle
validate_project
export_graph
Most MCP tools are read-only over the current index. load_remote_bundle mutates only the server's in-memory index by fetching a public GitHub tree; it does not write files. Concept listing and search default to compact summaries; pass detail: "full" when navigation metadata, signals, ranking, or snippets are required. Relationship paths are deduplicated by node sequence even when parallel edge kinds connect the same concepts.
Every MCP tool includes a purpose-specific description, descriptions for its input parameters, and standard annotations covering read-only behavior, destructive behavior, idempotency, and external access.
Tool arguments are validated against the advertised input schemas before
execution. Unsupported fields, missing required values, incorrect primitive
types, and out-of-range integers are rejected without coercion. Unknown or
disabled tool names remain protocol-level invalid-parameter errors.
Tool discovery and direct invocation use the same capability checks:
| Mode | Normal proposals | Direct live write | Computation proposal | Runtime remote load |
|---|
| default | disabled | disabled | disabled | disabled |
--authoring | enabled | disabled | disabled | disabled |
--write --actor openai/gpt-5.6 | disabled | enabled | disabled | disabled |
--authoring --allow-computation-authoring | enabled | disabled | enabled | disabled |
--allow-remote-tool | disabled | disabled | disabled | enabled |
An explicit local root or project workspace exposes concept validation, path suggestion, and proposal inspection helpers. Proposal mutation tools require --authoring. The direct live tools instead require --write plus a truthful actor using human:<id>, process:<id>, or provider/model syntax. In normal single-root mode, callers omit bundle; it is required only to select among multiple project roots. Remote roots remain read only.
Generic concept tools cannot create or change an Attested Computation contract. okf_propose_attested_computation additionally requires --allow-computation-authoring and creates one coordinated review proposal for the concept plus an optional external computation file.
Live Concept Authoring
Start the server with the direct-write capability only when the MCP client/user approval boundary is sufficient review:
okf --root /path/to/catalog --write --actor openai/gpt-5.6 mcp
The agent sees one read-only batch validator and one destructive apply tool. It supplies structured concept fields rather than YAML; OKF serializes compatible Markdown frontmatter and stamps the configured generated.by plus one generated.at timestamp for the whole batch.
{
"name": "okf_apply_changes",
"arguments": {
"message": "docs(okf): document order creation",
"changes": [
{
"op": "create",
"type": "MCP Tool",
"title": "Create Order",
"body": "# Create Order\n\nCreates a validated order.",
"tags": ["orders", "mcp"],
"sources": ["/repositories/orders-service.md"],
"relations": [
{ "type": "related_to", "target": "/workflows/order-creation.md" }
]
}
]
}
}
Create paths are optional. The server first uses a strong dominant directory convention from existing same-type concepts within the requested prefix, then falls back to deterministic type/title slugs. okf_suggest_concept_path reports the strategy, evidence, path availability, and same-type/title matches so an available filename is not mistaken for a safe duplicate. Updates identify an existing uri; stale locators return bounded likely replacements, while scalar fields replace existing values and tags, sources, and relations use explicit add/remove patches. metadata carries extension frontmatter but cannot override identity, generation, collection, or computation fields. Paths and URIs are immutable during update: moving a concept changes its portable identity and remains a separate, intentionally unsupported operation.
Every 1–100 item batch is validated as one future graph, so concepts created together can reference one another and same-type/title conflicts are detected across creates and updates. Call okf_validate_changes with the complete intended batch to receive a time-of-check preview without writing files. Validation and apply share the same planner; apply repeats every check under the writer queue because revisions and Git state can change after a preview. Compact receipts are the v0.8 default; pass detail: "full" for the v0.7 planning layout. Effects expose structured relations and keep server-managed generation provenance separate from substantive changedFields.
The server writes nothing unless every candidate is valid, revision checks still match, and every target stays inside one writable bundle. Process-generated documents, generator output directories, hidden/control-plane paths such as .git/**, reserved files, and Attested Computation contracts are not live-write targets. Rollback checks revisions immediately before each restore and reports detected replacements as a partial rollback_conflict. That protection is best effort under the documented single-external-writer requirement; it is not a cross-process compare-and-swap guarantee.
Add --git-commit as server policy to create one commit per successful batch. A detected Git worktree must be completely clean and have a configured identity before publication. Active Git filter attributes and assume-unchanged/skip-worktree index flags block the operation so validation cannot execute configured filters or overlook hidden user changes; replace-object resolution is disabled so hidden replacement history cannot alter the parent tree. The server builds an isolated index from the validated Markdown bytes, creates that exact tree with commit-tree, publishes it with a compare-and-swap ref update, synchronizes only the affected ordinary-index paths afterward, and never pushes. Concurrent unrelated staged entries cannot enter the commit. A determinate commit failure leaves matching valid files as working-tree-only; an ambiguous ref-update timeout is reported as unknown unless the resulting commit tree can be proven. Non-Git catalogs are written normally. Every response makes the repository root, commit state, index state, target-byte state, and persistence boundary explicit.
Write policy
A write policy narrows what each actor may write through live batches and proposals. It is trusted configuration: declare it as writePolicy in okf.project.yaml, or pass --policy <file> (the same mapping as a YAML file) with --root and --write or --authoring. Tool and HTTP arguments cannot name, replace, or relax it.
writePolicy:
defaultEffect: deny
rules:
- name: humans
actors: ["human:*"]
- name: nightly-seo
actors: ["process:nightly-*", "openai/*"]
bundles: ["customers-*"]
paths: ["findings/**"]
types: ["SEO Finding"]
requiredFields: ["sources", "stale_after"]
defaultStatus: draft
maxConceptsPerBatch: 20
- name: no-drafts
actors: ["**"]
paths: ["drafts/**"]
effect: deny
applyToProducers: false
For each concept in a batch, the first rule whose actors match the configured --actor and whose other selectors (when present) match decides. effect is allow by default; requiredFields, defaultStatus, and maxConceptsPerBatch apply to allow rules. Patterns use a fixed glob subset: * within one segment, ** as a whole segment for any depth, and literal segments. They are never compiled to regular expressions, and absolute, empty, . and .. segments are rejected when the policy loads. A malformed policy is a project error, so no writable server starts on it. The policy judges only the concepts a batch creates or updates; existing content is never retro-validated.
Violations are structured: { code: "policy_violation", rule, reason, field?, path?, bundleId?, uri?, limit?, count? }, where rule is the rule's name, else its index, and null when defaultEffect decided. reason is one of actor_denied, bundle_denied, path_denied, type_denied, missing_required_field, or batch_too_large. okf_validate_changes reports them read-only in policyViolations and as layer: "policy" diagnostics; okf_apply_changes refuses such a batch with status rejected before staging or touching any file. Receipts carry policyViolations only when a policy is configured; messages name rules, paths, and URIs, never catalog text.
Scope. Live batches (okf_validate_changes, okf_apply_changes) are judged against --actor. Proposals are judged twice. When okf_propose_concept, okf_propose_update, okf_propose_attested_computation, okf_propose_v02_migration (every child update and the index.md version declaration, before any is stored) or their HTTP equivalents would store one, it is judged against the proposing actor (--actor, which --authoring also accepts); a refused proposal returns created: false with policyViolations and is not stored. When okf_accept_proposal or POST /v1/proposals/<id>/accept would write it, it is judged against the actor recorded on the proposal (proposedBy), else the accepting actor; a refusal returns accepted: false before any file is written, and an accepted create the matching rule gives a defaultStatus is published with that status. A write with no actor (an HTTP call, or a proposal stored before an actor or a policy was configured) is judged as the literal actor unknown: * and ** match it and a rule can name it, so legacy proposals are judged as unknown at acceptance unless the accepting server has an actor. okf serve takes the policy from okf.project.yaml, or from --policy <file> with --root and --authoring. Producers publish under process:<producer name> and are exempt unless applyToProducers: true, in which case okf_preview_producer reports the violations of the created and updated concept files read-only and okf_run_producer / producer run refuse before publication. Producer output is never rewritten, so defaultStatus does not apply to it: a required status must be in the produced file, or the run is refused with missing_required_field. The reserved index.md and log.md files of a producer bundle are not judged.
Authoring Concepts
The reviewable proposal workflow remains available through MCP tools started with --authoring and through the HTTP API. Clients never need direct local file access.
MCP proposal flow:
{
"name": "okf_propose_concept",
"arguments": {
"path": "tools/create-order.md",
"frontmatter": {
"type": "MCP Tool",
"title": "Create Order",
"relations": [
{
"type": "related_to",
"target": "/workflows/order-creation.md"
}
]
},
"body": "# Create Order\n\nCreates an order through the application MCP tool.",
"message": "Document create_order for agents."
}
}
Then call okf_accept_proposal with the returned proposal.id.
To correct an existing concept, read it with get_concept, then propose only the fields that need to change:
{
"name": "okf_propose_update",
"arguments": {
"uri": "okf://app/tools/create-order",
"frontmatter": {
"title": "Create Order Tool",
"description": "Creates a validated order."
},
"removeFrontmatterKeys": ["deprecatedField"],
"message": "Correct outdated tool metadata."
}
}
Omitted frontmatter fields and an omitted body are preserved. The concept URI cannot change through an update. Each update proposal records the source file revision, and acceptance checks it again immediately before replacing the file so detected concurrent changes are rejected.
Safety rules:
- concept paths must be safe, non-hidden relative
.md paths inside a writable bundle
- concept writes cannot traverse symbolic links under a writable bundle
- missing subdirectories are created only when a proposal is accepted
index.md and log.md cannot be authored as concepts
- duplicate paths and duplicate
okf:// IDs are rejected
- updates cannot change concept identity and reject detected changes made after proposal creation
- invalid IDs, invalid relation types, and broken internal OKF relations fail validation
- external relation targets such as
repo://... are allowed
- direct batches are serialized within one process and validate the combined future graph before publication
- independent processes still require external single-writer coordination
HTTP API
The REST authoring API below is separate from MCP over HTTP (mcp --http).
Start the HTTP server:
OKF_WRITE_TOKEN=change-me okf --root /path/to/catalog serve --host 127.0.0.1 --port 8765
Read/validation endpoints:
GET /health
GET /v1/bundles
POST /v1/concepts/validate
POST /v1/concepts/suggest-path
Proposal inspection and mutation endpoints require Authorization: Bearer <OKF_WRITE_TOKEN> because pending records can contain complete candidate Markdown and computation code:
GET /v1/proposals
GET /v1/proposals/:id
POST /v1/proposals
POST /v1/proposals/update
POST /v1/proposals/:id/accept
POST /v1/proposals/:id/reject
The default file-backed proposal store writes proposal JSON under .okf-proposals in the selected root or project. Accepted proposals write Markdown concepts into the selected local root.
POST /v1/concepts/validate and POST /v1/concepts/suggest-path do not persist anything. POST /v1/proposals persists only a proposal record. Only POST /v1/proposals/:id/accept writes a concept Markdown file.
Remote Bundles
Remote bundles let one workspace consume concepts published by another repository without vendoring them. For a host-agnostic setup, clone or mount an OKF repository from any Git host and pass its directory through --root; transport and synchronization remain outside the OKF specification.
Supported source:
- Public GitHub repository tree URLs:
https://github.com/<owner>/<repo>/tree/<ref>/<path>
Remote loading:
- inventories the tree, fetches selected
.md documents first, then fetches only explicitly referenced bundle-local assets
- records resolved revision metadata, SHA256 digests, document/asset byte counts, and unresolved references
- inventories remote paths but never downloads the contents of unreferenced
.sql, .py, or binary files
- keeps each remote bundle under its configured bundle id
- supports
include and exclude filters
- resolves Markdown links inside the remote bundle path
- enforces file count and byte limits
- does not execute code from the remote repository
CLI examples:
okf --remote-bundle shared=https://github.com/example/okf-atlas/tree/main/bundles/shared --inspect
okf --project okf.project.yaml --remote-bundle vendor=https://github.com/example/vendor-okf/tree/main/bundles/catalog validate
MCP runtime loading:
Start the MCP server with --allow-remote-tool before calling load_remote_bundle.
{
"name": "load_remote_bundle",
"arguments": {
"id": "shared",
"url": "https://github.com/example/okf-atlas/tree/main/bundles/shared",
"include": ["public/**"]
}
}
Use list_remote_bundles to inspect what was loaded.
Structured Search
search_concepts accepts:
query
bundle
types
tagsAny
tagsAll
pathPrefix
frontmatter
linkedTo
linkedFrom
relationType
orphanOnly
statuses
trustTiers
freshness and deterministic asOf
hasSources
runtime and attestationReady
generatedBy and verifiedBy
detail (compact by default, or full)
limit
offset
list_concepts also accepts a text query (and mode) and applies it
together with its listing filters. Text search tokenizes case-insensitively,
folds diacritics, drops stop words and single letters, and stems every term
with the catalog's languages (English and Italian at 1.0). mode is auto
by default since 0.10: the concepts matching every term come first, in their
own order, and when fewer than autoFill of them exist (default 3) the
concepts matching only some of the terms follow them (the response's mode
is then auto; it is any when nothing matches every term and all when
nothing was appended). This keeps one long document that happens to contain
every term from hiding the shorter documents that match most of them.
autoFill is a library search option and an okf.project.yaml key
(autoFill: 0 restores the 0.10 rule: fall back only when nothing matches
every term), never a tool argument. all restores the pre-0.10
behaviour and any ranks anything matching one term. BM25+ ranks title,
type, tags, aliases, description, path, body, and an exact field (ids,
paths, aliases, raw title tokens) that is never stemmed, so identifiers and
SKUs still match literally. Double-quoted text is matched literally (no
stemming or accent folding, stop words and single letters kept) against ids,
paths, aliases, titles and raw tokens, and unquoted tokens containing digits
or hyphens stay literal as well; stems only ever reach the text fields.
Frontmatter remains available through exact structured filters but is not
copied into the text index. Scores are
relative within a result set and are not a stable cross-version scale; use
ordering and relevance instead.
Query responses report mode (the one that ran), queryTerms,
droppedTerms, allTermsTotal (concepts matching every term, also under
any) and lowConfidence, true when mode is any (the any-term ranking produced the page)
and its top result matched fewer than half of the terms. Each result carries
coverage (share of query terms matched), relevance (score over the top
score) and, in full detail, matchedTerms. Opt-in floors minCoverage and
minRelevance drop weaker results; freshnessPrior down-weights deprecated
(0.5), stale (0.7), and draft (0.85) concepts. Compact results contain
uri, title, type, description, coverage, and relevance; title,
type, and description are bounded, while full results remain lossless.
Queries are bounded to 512 characters and 16 terms after stop words. Prefix
expansion and fuzzy matching are intentionally disabled. The stemming
languages come from the project config key languages (auto, none, or
codes such as [it, en]) or --languages in --root mode; auto detects
each bundle's languages from its documents. Punctuation-only queries return
no matches. Tags and types are matched case-insensitively. Arbitrary
frontmatter filters support exact scalar matching and array-contains
matching. relationType selects concepts with an outgoing relation of that
type.
The SDK regression suite also budgets a neutral four-step research path from actual serialized MCP text. It uses UTF-8 bytes divided by four as a deterministic estimate, not an exact model tokenizer or billing count, and guards both an absolute compact budget and a compact/full ratio.
Example:
{
"query": "catalog",
"types": ["API Route"],
"tagsAll": ["api", "orders"],
"limit": 10
}
For local relevance and performance checks, run the non-packaged development
benchmark with a bundle root and an optional JSON array of { "query": "...", "expected": "path/or/concept-id" } judgments:
node --expose-gc scripts/search-benchmark.js \
--root /path/to/okf \
--qrels /path/to/qrels.json
It reports OKF and search-index build time, retained heap/RSS, p50/p95 query
latency, Recall@10, MRR@10, and representative rankings. Search indexes are
process-local and keyed to the parsed OKF index, so remote loads and accepted
proposals receive a fresh index automatically.
Budgeted context and section anchors (0.11)
get_context (MCP, read-only) and okf context <query> [--budget <tokens>] [--limit <n>] [--provenance compact|full|none] (CLI, JSON) answer a question
with the sections that matter instead of whole concepts. Candidates come from
the same search as search_concepts (same filters, mode, minCoverage,
minRelevance, lowConfidence, queryTerms). Their sections are scored on
their own text (BM25; the concept's type, tags and path add at most 20%) and
ordered by the candidate concept rank; within a concept, sections whose text
matches the query come before sections found only through that metadata, and
a strongly matching section may move ahead of up to four weaker concepts.
Sections are deduplicated and packed in that order while they fit
budgetTokens (default 1,200, never exceeded). The budget covers the whole
response as delivered: usedTokens is ceil(characters / 4) of the entire
JSON result, every field included (the candidate concepts, queryTerms,
the counters and every section field, scoring fields too), and get_context
returns that JSON compact, so the text an agent receives is what was
measured. Sections are packed into what the rest of the response leaves;
when even the fields around them do not fit, the candidate concept list is
shortened from its tail (omittedConcepts), and only a budget smaller than
the query echo and the field names can be exceeded (then no section and no
concept is returned). Each section's tokens is the same estimate for that
section, and the estimator is named in the response. Before 0.11.0-rc.2 only
the sections' content fields were counted, so responses ran at 1.6-1.7 times
the budget. Provenance is
compact by default (status, freshness, verified as a boolean,
sourceCount); provenance: "full" returns status, staleAfter,
freshness, trustTier, verified, generated and sources, and
provenance: "none" (or includeProvenance: false) spends the whole budget
on the sections. The library exposes
assembleContext(index, options), conceptSections(doc) and
sectionUnits(index).
Section anchors are an okf-mcp extension; OKF v0.2 defines none and nothing
in a bundle needs them. A concept is cut at its document-level headings: the
text before the first heading (with the title and description) is addressed
by the concept URI alone, every other section by uri#slug. The heading
text is first normalised (the line breaks of a setext or wrapped heading and
every whitespace run become one space, then trimmed); the slug then follows
GitHub's rule: lower-cased, every character other than a letter, a combining
mark, a digit, an underscore, a hyphen or a space removed, and each space
turned into a hyphen (so ## Consentito / Allowed gives
consentito--allowed, as on GitHub). A repeated heading takes -1, -2,
... in document order. A reader that drops the fragment still lands on the
concept. An anchor depends only on the heading texts up to its own: edits to
any section body, whitespace-only edits and edits to later sections leave it
unchanged, while renaming the heading, or adding a heading with the same text
before it, changes it.
Embeddings (0.11, opt-in)
Dense retrieval is off unless configured, and the default install makes no network calls. 0.11 ships the embedder interface, a zero-dependency client for the OpenAI embeddings API and compatible servers (Ollama, vLLM, LM Studio), a vector cache kept outside the bundle under $XDG_CACHE_HOME/okf-mcp/<sha256 of the root>/ (~/.cache/okf-mcp/ or %LOCALAPPDATA%\okf-mcp\), so a populated cache never dirties a Git worktree and can be deleted at any time, and hybrid search over it.
Enable it in okf.project.yaml (or, for --root mode, in a JSON or YAML file passed with --embedder-config <file>, which may also be a bare embedder block; --cache-dir <dir> overrides the cache location):
embedder:
type: openai
baseUrl: https://api.openai.com
model: text-embedding-3-small
apiKeyEnv: OPENAI_API_KEY
dimensions: 512
cacheDir: ../okf-vectors
fusion:
alpha: 0.4
normalization: tmm
embedUnits: concepts
Embedder settings are never accepted as tool arguments. At startup the server validates the embedder, discovers the dimension when it is not configured, opens the cache and embeds the missing units in the background; it never fails to start because of the embedder. search_concepts then fuses the lexical ranking with cosine similarity over the cache by convex combination of normalised scores: each side is scaled by its best score above a lower bound: by default the theoretical minimum of its scoring function (0 for the lexical score, -1 for cosine; tmm); lexical-tmm uses the theoretical minimum on the lexical side only and minmax the observed minimum on both, as in Bruch et al., arXiv:2210.11934 (hybrid: false per call keeps it lexical and makes no request). Remote bundles stay lexical-only: their text is never embedded. mode, coverage, allTermsTotal and the minCoverage/minRelevance floors keep their lexical meaning (an explicit mode: "all" only re-ranks the all-term matches): the lexical hits are fused with the 200 nearest concepts by cosine, a fixed candidate set so every page of a query paginates the same ranking, and each candidate carries its own lexical evidence (matched terms, coverage, score), which the floors and lowConfidence (computed on the displayed top result) read. When the all-term pass ran and the query holds quoted or identifier-like terms, the all-term matches stay ahead of every semantic neighbour. relevance and score become fused and each result adds lexicalScore, denseScore (cosine) and fusedScore. The response adds hybrid, alpha, normalization, embedder: { id, remote, dimension, degraded, code?, serverCode? }, cache: { rows, dir? } (dir only on a --write server), usage: { promptTokens, requests } for the query encoding, and timings: { queryEncodingMs, searchMs, totalMs }. Whenever the embedder is missing, failing or the cache is empty, search runs lexical-only with embedder.degraded: true, a fixed code and, for HTTP failures, the enumerated serverCode; no server-supplied text is ever returned. After a failed query encoding the server stays lexical for 30 seconds before trying again. After each successful okf_apply_changes batch only the changed units are embedded (one request when the batch fits) and the receipt carries embedding: { units, embedded, skipped, embedder, cache, usage, durationMs }. The read-only embedding_status tool, listed only when an embedder is configured, reports the embedder, the cache, the startup and last preparation receipts and the token usage. Receipts report embedder.remote: true whenever text leaves the machine. The cache is keyed by the embedder identity openai:<endpoint>:<model>:<dimension>:<maxInputChars>, so a change of endpoint, model, dimension or truncation length rebuilds it. The library exposes createOpenAiEmbedder, openVectorCache, resolveCacheDir, prepareEmbeddings, openEmbeddingRuntime, denseSearch and fuseRankings; node scripts/eval-search.js --embedder <file> scores hybrid runs on the retrieval fixtures.
Long or identifier-dense documents: the client truncates every input at maxInputChars (default 24,000 characters) and plans requests with a conservative estimate of one token per 3 characters, because generated contracts and column lists run at about 2.6 characters per token and overflow the provider's 8,192-token input limit at the old 32,000-character default (the default change rebuilds caches keyed with 32000). If the provider still refuses a request as invalid_request, the client splits the batch in halves until the refused inputs stand alone (a bounded number of extra requests), retries each refused input once at half its length (never below 2,000 characters) and, if it is refused again, skips that unit: the preparation receipt lists it under skipped: [{ id, reason: "input_rejected" }], every other unit is embedded and search stays hybrid. embedding_status reports skipped: { count, ids } (the first five ids) for the last preparation, and the startup receipt carries the same field; a skipped unit is retried by the next preparation and is found by its lexical score meanwhile. A request whose every input is refused is a configuration problem and still fails as before.
Adaptive fusion weight (0.12): fusion.alpha: adaptive decides the weight of the semantic side once per query, from what the lexical pass already knows, so every page of a query uses the same weight: fusion.adaptive.low when the query holds quoted or identifier-like terms and some concept matches every term (the all-term matches still stay ahead of every semantic neighbour), mid when at least autoFill concepts match every term, high when no concept matches every term and none matches even half of them, and in between linear in the best lexical coverage, from high at 0.5 to mid at 1. The weights default to the development-split choice { low: 0.1, mid: 0.5, high: 0.9 }; high may not be below mid. Hybrid responses report alphaUsed and alphaReason (fixed, literal_terms, all_terms, weak_lexical or interpolated) and embedding_status reports the setting and its weights. The default stays the fixed 0.4, which ranks exactly as 0.11 did: on the fixtures the adaptive weight raised alpinaria Recall@5 from 0.830 to 0.838 (synonyms 0.542 to 0.604) but lowered atlas from 0.882 to 0.870, so it missed its pre-registered conditions (test/fixtures/retrieval/reports/hybrid-0.12-2026-10-25.md).
fusion:
alpha: adaptive
adaptive: { low: 0.1, mid: 0.5, high: 0.9 }
Contextual units (0.12): embedUnits: contextual-sections embeds each section after a one-line prefix naming its concept, title · type · path (at most 160 characters; a long title is shortened first), so a bare "Limits" or "Steps" section carries the system it belongs to; contextual-concepts puts type · path before the concept text. Anchors are those of sections. The prefixed texts have their own hashes, so switching unit sets re-embeds every unit once (the preparation receipt, the startup receipt and embedding_status name the unit set; embedded and pruned count the switch), and afterwards a body edit re-embeds only its section (a title, type or path edit re-embeds every section of that concept). concepts stays the default: on the development queries no other unit set beat it by the pre-registered rule (mean dev Recall@5 0.844 for concepts, 0.843 contextual-concepts, 0.841 contextual-sections, 0.834 sections), and on the fixtures' short sections the prefix costs 56% more prompt tokens (6.6% for contextual-concepts).
Confidence (0.12): every search_concepts and get_context response to a query carries confidence: { coverage, topSimilarity, lowConfidence, reasons }. coverage is the share of query terms the top result matched; topSimilarity is the best cosine similarity between the query and any candidate's units when the semantic side ran (null otherwise); lowConfidence (also the top-level field) is true when the any-term fallback produced the results and the top result matched fewer than half of the terms (reason weak_lexical_match), or when the top result matched fewer than half of the terms and topSimilarity is below confidence.minSimilarity (reason low_similarity). Lexical-only responses have no similarity and keep the 0.10 rule. get_context takes abstainOnLowConfidence: true to return no sections for a low-confidence query (the response keeps the confidence block and the candidate concepts and says abstained: true). The similarity threshold depends on the embedding model; the default, 0, was chosen on the development queries for text-embedding-3-small: no threshold up to 0.8 flagged a single extra unanswerable query there, because every unanswerable query whose top result matched fewer than half of the terms was already an any-term fallback. A deployment can set its own:
confidence:
minSimilarity: 0.3
Graph Behavior
As an okf-mcp graph projection, Markdown links become markdown_link edges, extension relations become typed relation edges, and standard v0.2 path-valued fields become resource, source, computation, executor, and attester edges. Internal concept references resolve to canonical nodes; explicitly referenced non-Markdown files resolve to okf-mcp okf-asset:// nodes; URLs and scope descriptors remain unfetched external or opaque leaves.
For navigation convenience, okf-mcp resolves links to a nested bundle directory to that directory's reserved
index.md when there is no exact document target. This applies to local and
remote bundles and to candidate validation during proposal authoring.
Graph tools return bounded JSON:
{
"nodes": [
{
"id": "okf://app/routes/order-status",
"bundle": "app",
"path": "routes/order-status.md",
"type": "API Route",
"title": "Order Status Route",
"tags": ["api", "orders"],
"description": "Serves order status state."
}
],
"edges": [],
"warnings": []
}
Use graph_summary first for counts by lifecycle, trust, freshness, runtime, readiness, and edge kind. Graph tools accept edgeKinds; pass includeExternal: true or includeAssets: true when those leaf nodes are needed.
Default relation types:
depends_on
produces
consumes
persists_to
materializes_to
configured_by
checked_by
owned_by
supersedes
related_to
Add project-specific relation types with relationTypes in okf.project.yaml.
Project paths in bundles and plugins must be relative paths that stay inside the directory containing okf.project.yaml. Absolute paths and ../ escapes are rejected.
Bundle include and exclude filters use simple path patterns:
- exact file paths, such as
services/order-status.md
- directory prefixes, such as
archive/
* for one path segment
** for any nested path
Validation
validate, validate_bundle, and validate_project return separate conformant and validForProject fields plus structured diagnostics. valid remains a compatibility alias for validForProject.
OKF conformance covers, when the corresponding files are present:
- parseable YAML mapping frontmatter on non-reserved concept documents
- a non-empty
type
- the reserved structure of
index.md and log.md
Unknown frontmatter keys and unknown concept type values do not fail conformance. The YAML parser supports nested mappings, arrays, block scalars, and other structures accepted by its safe YAML core schema; duplicate keys and unsupported custom tags are rejected.
Missing index.md files and broken cross-links do not fail OKF conformance. strictLinks affects only okf-mcp workspace validity (validForProject), not the normative conformant result.
Project validity additionally reports:
- duplicate OKF URIs
- broken internal Markdown links as advisories by default; set project
strictLinks: true or pass --strict-links to make them project-invalid
- invalid relation types
- missing relation targets
- broken
okf:// relation targets
- duplicate bundle IDs
- invalid or escaping project paths
- links that resolve outside the configured bundle root
- missing bundle roots
The server keeps serving valid concepts from partial bundles.
Optional v0.2 families are normalized into signals. Malformed provenance, generation, verification, lifecycle, freshness, or computation metadata produces an advisory and never creates a fourth trust tier. Verification fails closed to unverified; absent status defaults to stable; freshness is evaluated at an explicit asOf date when supplied. Authoring is stricter than consumption and rejects malformed known v0.2 fields.
Attested Computation
inspect_attested_computation reports the runtime, declared parameters, sanctioned inline or file computation digest, executor receipt fields, attester reference, indexed assets, readiness, and diagnostics. prepare_attested_computation checks declared parameter names and returns digests without returning values. check_computation_receipt checks field presence without returning values, persisting the receipt, or claiming attestation.
okf-mcp has no execution or attestation adapter. It never runs the computation, executor resource, or attester resource, and it never fetches an external contract URI on demand.
CLI parity is available through computation inspect|prepare|check-receipt, provenance, edge-kinds, and asset. Supply sensitive values with --parameters-file <path|-> or --receipt-file <path|->; raw parameter and receipt JSON is intentionally rejected in process arguments. - reads one JSON object from stdin. Asset reads accept --max-content-bytes up to the indexed 1 MiB limit.
Migrating Existing Catalogs To v0.2
The v0.2 specification keeps v0.1 bundles consumable through two fallbacks: legacy timestamp when generated is absent, and a legacy body # Citations list when sources is absent. okf-mcp applies those fallbacks during reads and offers an optional review-only conversion workflow.
For one root, inspect migration readiness and preview the proposed native fields without writing anything:
okf --root /path/to/catalog migrate check
okf --root /path/to/catalog migrate preview \
'{"metrics/revenue.md":{"by":"human:owner","confirmed":true}}'
In optional multi-root project mode, supply the root id before the actor-mapping JSON.
Migration is deliberately conservative:
- native
generated and sources fields always win
- a valid
timestamp is copied into a new generated: { by, at } mapping only after a truthful by actor is explicitly confirmed
# Citations becomes sources only from one top-level H1 section containing at least one safely parseable list entry and no unparsed prose, nested sections, ambiguous entries, or escaping paths
- legacy fields and citation prose are retained for compatibility
- concepts marked by
--generated-path, a document flag, or generated_file/generatedFile frontmatter must be changed through their generator; remote roots are report-only
- identity collisions, invalid documents, unsafe references, and unresolved assets block the version declaration
okf_propose_v02_migration requires local-root authoring. It creates a review manifest, one proposal per affected file, and a gated root okf_version: "0.2" proposal. Nothing is accepted automatically; the root proposal can be accepted only after every child is accepted and the complete catalog validates.
Generator Plugins
Generator plugins are configured in okf.project.yaml and run with generate.
Built-in plugins:
filesystem: creates one concept per matching source file. Defaults to Markdown files.
json-spec: creates one concept per JSON file and can emit persists_to relations when a destination table is present.
Generated output is regular Markdown/YAML OKF and is validated by the same indexer as hand-authored concepts.
Limitations
- MCP is served over stdio and, with
mcp --http, over Streamable HTTP. The separate HTTP authoring API (serve) is not MCP over HTTP.
- MCP protocol compatibility follows the pinned official SDK v2 dependency.
- There is no file watcher. Over stdio, restart the server after external file changes; over HTTP, changes are picked up before a request at most once per
--refresh-interval. Concepts accepted through MCP authoring refresh the MCP server index immediately.
- With
mcp --http, proposal authoring (--authoring) is not available, a revoked introspected token stays valid for up to 60 seconds, and one process still serves one root or project for every principal.
- The HTTP API is a lightweight built-in server, not a full hosted multi-tenant service.
- OKF v0.2 computation support is static inspection and preflight only; no computation or attester is executed.