Osnova

website getosnova.dev

A deterministic code map for AI coding agents. Osnova indexes a repository into a symbol and call graph with tree-sitter, then serves it to any MCP client or from the command line. Same input, same output, byte for byte. No embeddings, no telemetry, and no network connection unless you type osnova update-check. Twenty languages, seven of them (TypeScript, JavaScript, Python, Go, Rust, Java, C#) with deep adapters.

The getosnova.dev film in eight scenes, measured on osnova 0.12.0 indexing its own source: the word osnova; 442 files read into 7,896 symbols and 35,343 edges; one function parsed into its tree-sitter tree; eleven of the calls refreshWorkspace makes, woven as threads; an agent asks who calls refreshWorkspace and osnova_warp answers with exact file and line; two fresh builds give identical hashes; the install command.

Osnova is the Slavic word for base or foundation. That is the job: give an agent solid ground to stand on before it edits code.

Quick start

Node.js 22.13 or newer. Serve a repository to any MCP client:

npx -y @getdomovoi/osnova mcp --workspace /path/to/repo

After npm install -g @getdomovoi/osnova, one command per kind of agent: claude wires Claude Code (MCP entry, session, prompt and stop hooks, and the skill); agents wires every installed harness that reads AGENTS.md (Codex, OpenCode, Kilo, Pi, Cursor), each with its hooks, plugin or extension, plus one shared skill in ~/.agents/skills/. Each previews its changes; --apply writes them, backing up every file it changes.

osnova setup claude --apply
osnova setup agents --apply

Or add the MCP entry by hand; replace osnova with npx -y @getdomovoi/osnova when there is no global install.

{ "mcpServers": { "osnova": { "command": "osnova", "args": ["mcp"] } } }

What you get back

osnova warp src/api.ts#refreshWorkspace on this repository, cut to twelve lines (test/docs-readme-warp.test.ts fails when the first two no longer reproduce):

function src/api.ts#refreshWorkspace: 83 indexed edges
reach: d1 callers 83 in 14 files (3 dirs); d2 +16 in 2 files; unresolved same-name 10; tests 18
d1 calls src/cli/cli.ts#ensureIndex:78 [import-binding]
d1 calls src/cli/hook.ts#runHook:191,203,225,240 [import-binding]
d1 calls src/mcp/server.ts#createOsnovaMcpServer.refresh:297 [import-binding]
d1 calls test/verification-fastpath.test.ts#<module>:37,47,48,51,53,57,58,64,68,74,78,85,110,114,132,145,154,166,171,197 [re-export-binding]
  via src/index.ts:65 refreshWorkspace -> src/api.ts (export refreshWorkspace)
This does not prove absence of callers or that deletion is safe.
unresolved evidence (10); not confirmed relationships
candidates for refreshWorkspace (2, unverified): src/api.ts#refreshWorkspace, test/mcp-watch.test.ts#refreshWorkspace
d1 calls refreshWorkspace test/artifact-format9.test.ts:78,85,104,120,140 [binding-blocked]
d1 calls refreshWorkspace scripts/perf.mjs:485,486,491 [import-target-unresolved]

Each confirmed line names the caller, its lines and the evidence that tied the call to this definition: an import binding, a same-file definition, a re-export chain with its hop, or an identified receiver. Calls the index could not tie to a definition are listed apart as unresolved evidence, with the same-name candidates it found and the reason it stopped, so a name match is never mistaken for a caller. The reach line gives exact counts, not scores, and when the list outgrows its budget a capped: line and an omitted: footer count what was left out.

How much of the graph is exact

A call site counts as resolved when the index ties it to one definition through evidence it can name; everything else stays unresolved with a reason. These are the shares on the pinned checkouts under benchmarks/corpora/, recorded in benchmarks/results/resolution-coverage-2026-09-21b.json; the last column leaves out calls through packages outside the repository and calls to builtins, which can never resolve locally, and the reference defines every column.

Per-language rows are in the reference. osnova coverage reports the same numbers for your own repository, per language and per reason.

Resolved is not the same as right, so the call edges are also scored against each language's type checker or compiler. Every call site in six pinned checkouts was sent to the checker for the callee's declarations, and each osnova edge was marked true when the definition it names contains that declaration and false when it does not. Measured at 0.12.0:

Java and C# overloads share one symbol in the index, so a call edge names the overload it binds only when the written argument count fits exactly one declaration of the type, its partial parts and its declared base classes; when it fits several or none, the edge keeps the method and names no declaration, and the oracle leaves it out of both rates (2526 edges on gson, 602 on humanizer). 0.11.0 named the last overload declared for every call, recorded a call at the first line of its call expression, and lost the rest of a C# file after a primary constructor or a raw string; it measured gson 2045 false edges (24.75%), humanizer 617 (8.25%) and ripgrep 112 (2.06%). Of the 29 left on ripgrep, 11 are functions declared once per #[cfg] branch, 6 are a test module's function named like its parent's, and 5 are calls to a local closure or parameter named like a function elsewhere. A call whose target has one declaration that fits keeps it even when a base type is outside the index, where an unseen base overload could be the one the compiler binds; this limit is kept because withdrawing those edges cost 10 recall points on humanizer for one fewer false edge. Every false edge is classified by cause in benchmarks/results/type-checker-oracle-2026-09-30.json with the method and its limits. The scripts that produce these numbers are in benchmarks/oracle/.

Grep versus the graph

The reason to keep a call graph instead of running a text search is not speed. It is that the first regex a person types is wrong more often than it looks, and nobody notices. Nine call-site sets in six languages were verified line by line; each cell shows sites found (precision / recall), and the method lists what each side got wrong.

The graph never returned a site that was not a call of the target. The record is benchmarks/results/grep-vs-graph-2026-09-21.json.

The ten tools

The names play on the foundation image. The CLI uses the same names without the prefix: osnova ground and osnova_ground are the same query.

Every MCP response opens with its index generation, says when the index is partial, and counts what its budget left out; the budgets, and which CLI subcommands carry the generation, are in the reference.

Hooks and clients

One global install serves every repository and every client: one entry in each client's global config, nothing per project, nothing written inside your repository. osnova setup claude and osnova setup agents show the diff and write nothing until --apply; agents skips a harness whose config folder is missing, --only codex,pi narrows it, and a skill or plugin file you edited is kept and reported. --uninstall reverses either family the same way, previewing first. The table lists the per-client form, for one piece at a time. What each hook prints, when it stays quiet, and what the trials measured are in the reference.

In CI

The action runs osnova settle --base-ref against the pull request base and lists every indexed dependent of the symbols the pull request changed; the report is indexed structural evidence only, so a missing dependent is not proof that nothing depends on the change.

- uses: actions/checkout@v4
  with:
    fetch-depth: 0
- uses: getdomovoi/osnova@v0.12.0
  with:
    depth: "2"

Inputs and the local form are in the reference.

What osnova does not do

• No type inference and no dynamic dispatch. Edges come from syntax: direct calls, imports, name references, declared heritage (a written superclass or interface name) and framework routes (a registration whose receiver binds to a listed framework import, with the verb and the literal path written at the site), with lexical binding and receiver hints for TypeScript, JavaScript and Python. Resolution is heuristic and says so.
• No route table. A routes edge is one registration site tied to one handler; prefixes from mounts, blueprints and controllers are recorded on their own edges and never composed into a full path, a computed path records no path, and an inline closure or a wrapped handler records the route with no target. Express, NestJS, Flask and FastAPI are read; gin, axum, Django, Spring, ASP.NET, Rails and file-based routers are not.
• That boundary has a measured price. Scored against a type checker on two pinned corpora, the calls osnova does not resolve are mostly calls whose receiver type is never written down: 2945 of 6359 missed sites on zod and 164 of 307 on click are a plain name carrying no annotation, and another 1478 on zod are a call result or a property chain. Only 349 missed sites on zod and 10 on click have a type written at the receiver's declaration, and 248 of those 349 are a single library idiom. Resolving every one of them would move recall from 76.9% to 78.2% on zod and from 90.4% to 90.7% on click, so the boundary costs roughly one recall point rather than ten. The full census is in benchmarks/results/receiver-boundary-census-2026-09-21.json.
• No semantic search. osnova_ground is fielded lexical ranking over definitions. It is fast, deterministic and explainable, and it will not match a paraphrase.
• No proof of safety. An empty caller list means the index found no caller, not that none exists.
• No cost claims. Agent trials so far show correctness parity with and without the graph on small tasks. A benchmark that separates the two is in progress.

How it stays honest

• Every query refreshes the index from the working tree first, uncommitted edits included, and reports its generation.
• Every clipped output says how much was clipped. Every ranked list says how many candidates it dropped. Partial indexes say so on every response.
• Every hit carries a file:line span, a source hash and an index generation, so you can check what the agent cites.
• Every benchmark result in benchmarks/results/ is frozen with its corpus fingerprint, and rejected experiments stay on record next to accepted ones.
• The cache verifies a SHA-256 over the structural core before parsing it and hash-checks each source text on read. Nothing is written outside it.

CLI

Every tool is a subcommand: osnova build <root>, osnova warp <symbol>, osnova settle --base-ref <ref>, plus coverage, check, doctor, setup and mcp. The full list is in the reference.

Library

import { buildIndex, ask, callersDetailed, renderMapCard } from "@getdomovoi/osnova". The structured APIs return complete results with omission counts; the text budgets apply to CLI and MCP presentation only. Every export is in the reference.

Contributing

Bug reports, language adapters, benchmark corpora and agent trials are all welcome. Read CONTRIBUTING.md for the gates a change must pass: lint, typecheck, tests, build, perf budgets and package smoke.

pnpm install
pnpm lint && pnpm typecheck && pnpm test && pnpm build && pnpm perf

Determinism is the core invariant. A change that makes incremental refresh differ from a full rebuild by one byte is a bug.

License

Apache-2.0

Privacy: PRIVACY.md. Security policy and reporting: SECURITY.md.