@adrkit/mcp

Deterministic, offline, read-only decision memory for coding agents — including the decisions you already rejected.

A local Model Context Protocol server that lets an agent harness query one Git-backed ADR (Architecture Decision Record) corpus over stdio: "has this been decided?", "what governs these files?", and "what replaced this?" It surfaces superseded and rejected decisions specifically so agents stop re-proposing paths the team already ruled out.

No model calls. No network calls. No writes. The server reads Markdown files from your repo and returns structured JSON. It never calls an LLM, never opens a socket, and never mutates your corpus. Every tool is annotated readOnlyHint: true, openWorldHint: false.

Part of adrkit. The corpus lives in git as one Markdown file per decision with typed YAML frontmatter (@adrkit/core); this server only reads it. Listed in the official MCP registry as dev.adrkit/mcp (this is the Node/npm @adrkit/mcp package — unrelated to the adr-kit Python package on PyPI).

Status

Published on npm for Node 22+ with a deliberately small, read-only stdio surface: four tools and no write path.

adrkit is still pre-1.0. If you need repeatable client installs, pin an explicit @adrkit/mcp version in your MCP config instead of floating latest.

Quick start

The published binary is adrkit-mcp. Run it with npx (no install):

npx -y @adrkit/mcp --cwd /path/to/your/repo --dir docs/adr

It speaks JSON-RPC over stdio, so you normally point an MCP client at it rather than running it by hand. Copy-pasteable client configs follow.

Claude Desktop

Edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json). Claude launches servers from an arbitrary working directory, so set ADRKIT_MCP_CWD to your repo's absolute path:

{
  "mcpServers": {
    "adrkit": {
      "command": "npx",
      "args": ["-y", "@adrkit/mcp"],
      "env": {
        "ADRKIT_MCP_CWD": "/absolute/path/to/your/repo",
        "ADRKIT_MCP_DIR": "docs/adr"
      }
    }
  }
}

VS Code

Create .vscode/mcp.json in your workspace (VS Code uses the servers key and substitutes ${workspaceFolder}):

{
  "servers": {
    "adrkit": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@adrkit/mcp", "--cwd", "${workspaceFolder}", "--dir", "docs/adr"]
    }
  }
}

Cursor

Create .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "adrkit": {
      "command": "npx",
      "args": ["-y", "@adrkit/mcp"],
      "env": {
        "ADRKIT_MCP_CWD": "/absolute/path/to/your/repo",
        "ADRKIT_MCP_DIR": "docs/adr"
      }
    }
  }
}

GitHub Copilot CLI

Add to ~/.copilot/mcp-config.json (user-level) or ./.copilot/mcp-config.json (per-repo). Copilot CLI runs servers from the trusted repo directory, so the default cwd usually resolves correctly:

{
  "mcpServers": {
    "adrkit": {
      "type": "local",
      "command": "npx",
      "args": ["-y", "@adrkit/mcp", "--dir", "docs/adr"],
      "tools": ["*"]
    }
  }
}

You can also add it interactively with /mcp add inside a copilot session.

Configuration reference

Flags win over environment variables, which win over the defaults. An unusable configuration exits non-zero with a diagnostic on stderr (2 for an unparseable flag, 1 for an invalid root/directory) and never starts a transport. stdout is reserved for JSON-RPC protocol frames only.

Protocol revisions

The server speaks both MCP protocol eras on the same stdio connection, and the client picks. The opening exchange selects the era and pins it for the connection's lifetime:

On 2026-07-28 there is no initialize handshake and no session id: every request carries its own protocol version and client capabilities in _meta, and every result is self-describing (resultType, plus server identity in _meta). tools/list and server/discover are cacheable (SEP-2549) and are served with ttlMs: 300000, cacheScope: "public" because the 4-tool surface is fixed for the life of the process and carries no corpus content. Corpus reads are never cacheable: every tools/call loads a fresh projection.

Nothing else about the tools changes between eras — same names, same schemas, same annotations, same structured results. The server uses none of the features the 2026-07-28 revision deprecated (roots, sampling, logging) or removed (sessions, ping, resources/subscribe).

The four tools

Exactly four tools, all read-only. Each shares fixed annotations (readOnlyHint: true, destructiveHint: false, idempotentHint: true, openWorldHint: false), returns a deterministic human-readable summary line in content[0].text, and carries a findings page with the corpus's own parse/validation findings. Every substantive response also includes a corpusHealth sibling: { fingerprint, recordCount, excludedCount }.

Shapes below are the real input/output contracts (see packages/mcp/src/tools). Each tool's structured output is { corpusHealth?, result }; the objects shown under "output" are the members of that discriminated result union (keyed on outcome). findings and every growing array are cursor-paginated (see Pagination); findings and repeated pagination fields are elided here for brevity.

search_decisions

Normalized literal substring search over id, title, tags, and body (the graveyard of superseded/rejected records is included by default). Filters are ANDed: status/scope match any-of; tags matches all-of.

// input
{
  "query": "postgres",          // required, 1–256 code units, non-empty after trim
  "status": ["accepted"],        // optional, ≤6, any-of
  "tags": ["database"],          // optional, ≤32 tags × ≤64 chars, all-of
  "scope": ["backend"]           // optional, ≤3, any-of
}
// output → single "results" branch (empty results use the same branch)
{
  "outcome": "results",
  "items": [
    {
      "id": "0007",
      "title": "Adopt Postgres",
      "status": "accepted",
      "sourcePath": "docs/adr/0007-adopt-postgres.md",
      "matchedFields": ["title", "body"]   // subset of id | title | tag | body
    }
  ],
  "cursor": null
}

get_decision

The complete typed frontmatter + body for one ref.

// input
{ "ref": "0007" }               // AdrRef, 1–128 chars (local id, or a "log:id" federated ref)
// output → discriminated on "outcome"
// found (the full record is nested under "decision"):
{
  "outcome": "found",
  "decision": {
    "requestedRef": "0007",
    "id": "0007",
    "title": "Adopt Postgres",
    "status": "accepted",
    "sourcePath": "docs/adr/0007-adopt-postgres.md",
    "frontmatter": { /* typed YAML: status, tags, affects, relations, ... */ },
    "body": "## Context\n..."
  }
}
// other outcomes:
// { "outcome": "not-found", "requestedRef": "9999" }
// { "outcome": "ambiguous-local-id", "requestedRef": "0007", "candidates": [ /* DecisionSummary[] */ ] }
// { "outcome": "federated-log-unavailable", "requestedRef": "core:12", "log": "core", "id": "12" }

A log:id federated ref is recognized but never resolved or substituted — this server reads exactly one local corpus.

get_decision_context

Governing / active-proposal / historical decisions for repo-relative files[], via each record's own affects matchers. Paths are compared against patterns only — never opened.

// input
{
  "files": ["src/db/pool.ts", "src/db/schema.sql"]
  // 1–256 entries; each POSIX, 1–1024 chars; no leading "/", no "..", no drive, no "\"
}
// output → single "matches" branch (all three arrays; empty is the same branch)
{
  "outcome": "matches",
  "governing": [
    {
      "id": "0007",
      "title": "Adopt Postgres",
      "status": "accepted",
      "sourcePath": "docs/adr/0007-adopt-postgres.md",
      "firedMatchers": [ { "type": "path", "pattern": "src/db/**" } ],
      "relations": { "supersedes": [], "supersededBy": null, "relatesTo": [], "conflictsWith": [] }
    }
  ],
  "activeProposals": [],
  "history": []
}

list_superseded

Every superseded record with its direct local replacement state — the tool that keeps agents from re-proposing rejected paths.

// input (pagination only)
{}
// output → single "entries" branch
{
  "outcome": "entries",
  "items": [
    {
      "id": "0003",
      "title": "Use MySQL",
      "status": "superseded",
      "sourcePath": "docs/adr/0003-use-mysql.md",
      "supersededBy": {            // one of four states:
        "resolved": true,
        "target": { "id": "0007", "title": "Adopt Postgres", "status": "accepted", "sourcePath": "docs/adr/0007-adopt-postgres.md" }
      }
      // unresolved states:
      //   { "resolved": false, "targetRef": "0099", "reason": "dangling" }
      //   { "resolved": false, "targetRef": "0007", "reason": "ambiguous", "candidateCount": 2 }
      //   { "resolved": false, "targetRef": "core:1", "reason": "federated-unavailable", "log": "core", "id": "1" }
    }
  ],
  "cursor": null
}

Relation refs (supersedes, supersededBy, relatesTo, conflictsWith) are surfaced verbatim and never expanded — follow them with a second get_decision call. Supersession is reported one hop deep; there is no transitive traversal.

Library surface

The package root exports only a sealed lifecycle factory. There is no way to reach the underlying SDK server, its registrations, or its transport:

import { createAdrkitMcpServer } from '@adrkit/mcp';

const server = createAdrkitMcpServer({ cwd: process.cwd(), dir: 'docs/adr' });
await server.start();   // validates the root, then serves one stdio connection
// ... later:
await server.close();

createAdrkitMcpServer(options?) performs no filesystem access at construction and returns a frozen, null-prototype handle with exactly start() and close().

Transport failures — a transport that fails to start, or a background stream error such as the EPIPE from a client that has gone away — are reported through the optional onError option, which defaults to a stderr diagnostic. The adrkit-mcp binary additionally exits non-zero. Nothing is written to stdout: that is reserved for protocol frames.

Limits

query 1–256 code units (non-empty after trimming); ref 1–128; files[] 1–256 entries of 1–1024 POSIX-only chars (no leading /, no .., no drive letter, no backslash); status ≤6, scope ≤3, tags ≤32 × ≤64 chars; result and findings pages default 20, max 100; a per-record ADR source cap of 64 KiB (oversized records are excluded and surfaced as a record-too-large finding, never truncated). Every input object is strict — an unknown field is rejected before any corpus access.

Pagination and cursor restart

Every growing channel returns a cursor (null on the last page). Cursors are opaque, versioned, and bound to both the corpus fingerprint and the call's query shape. Reuse a returned cursor only with identical request parameters against an unchanged corpus; otherwise the response is a non-error invalid-cursor outcome (corpus-changed, query-mismatch, wrong-channel, offset-out-of-range, cursor-not-applicable, version-unsupported, or decode-failed) and you should restart the walk from no cursor. The primary-result and findings channels page independently. A corpus-unavailable outcome is returned when the ADR directory cannot be read.

Boundaries (out of scope by design)

No fifth tool; no writes, proposals, or PR creation; no MCP prompts, resources, subscriptions, or sampling; no HTTP/SSE transport or authentication; no model, embedding, ranking, or network access; no persistent cache/index/database; no named-log federation or multi-repository aggregation; no transitive supersession traversal.

Toolchain

Developed and tested with Bun; published artifacts are ESM targeting Node.js >=22 and are verified on Node 22 and 24. Runtime dependencies are exactly @adrkit/core, @modelcontextprotocol/server (MCP TypeScript SDK v2), and zod.

Apache-2.0.