retrieval-mcp

Package release: 0.4.11 (2026-09-15), 33 tools. Platform architecture: ../overview.md.

0.4.11 adds Retrieval-first initialize guidance and permits one MCP call to carry multiple independent query strings for search_code, journal_search, search_public_code, search_papers, and search_within_paper. A string remains fully compatible; a list preserves input order (including duplicates), runs each query with the existing filters/depth/rerank/formatter, and labels each result. The batch has bounded in-flight concurrency only; it never shares or trims candidates. 0.4.10 fixes a cross-project scope bug. The daemon is an account-level singleton, but a relative scope (scope_dir=".", repo_root=".", save_dir=".") was resolved against the DAEMON's own working directory — whichever project's shim started it. So journal reads returned a foreign repo, journal_record wrote into that repo's namespace, and downloads landed in the wrong project. The shim now sends its directory in the authenticated raw handshake and the daemon binds it per session with a context variable, so concurrent sessions from different projects stay independent. DAEMON_PROTOCOL_SCHEMA_VERSION goes 1 → 2 so a fixed shim cannot reuse an unfixed daemon. Tool surface is unchanged (33 tools, same digest).

Two changes in 0.4.9 over 0.4.8:

• index_code and journal_index_dir document the nested-scope guard's gitignore exception — a child scope indexing with respect_gitignore=False under a filtered parent now warns instead of being refused.

• Request bodies over 64 KiB are compressed with zstd level 3, which is what index_code/index_dir uploads are (a real 19.3 MiB body becomes 2.2 MiB in 54 ms). This adds a zstandard dependency; it also makes httpx advertise zstd, so replies come back compressed with no further work.

  It does not require a backend upgrade. Compression is negotiated per request: the client learns a backend accepts compressed bodies from the Accept-Encoding response header (RFC 7694), and a backend that does not rejects the body with 400/415/422 — all decided before the route function runs, so the automatic uncompressed resend cannot duplicate an index submission. After one such rejection the client stops compressing for the session.

find_duplicate_code was added in 0.4.8 and its backend endpoint is live on .88.

The package declares mcp>=1.29,<1.30. In addition to the public SDK v1 FastMCP API, the raw JSONL daemon bridge uses the verified 1.29 low-level SessionMessage and server-stream APIs. The narrow upper bound is therefore a private-API compatibility contract, not merely a v2 exclusion; raising it needs the daemon concurrency and clean-wheel MCP regressions first.

The daemon's 33-tool compatibility digest covers each tool name, full inputSchema, and semantic description text. Description canonicalization uses inspect.cleandoc only to remove the shared docstring margin emitted differently by the supported Python 3.10/Pydantic and Windows renderers; it does not reduce the fence to tool names or discard meaningful relative indentation.

Code indexing deliberately declares chonkie[code]>=1.6,<1.7, tree-sitter-language-pack>=1.8,<1.13, and tree-sitter>=0.23,<0.26. The local CodeChunker path is verified on Chonkie 1.6.x, the 1.12.x language-pack line, and the native tree-sitter 0.25.x ABI. Chonkie's code extra only gives its parser dependencies lower bounds; the explicit companion constraints prevent a fresh install from silently crossing into a newer parser-pack or native ABI line. The native constraint is intentional because local indexing executes native parser code and its crash boundary must not drift. Raising any upper bound requires a CodeChunker regression and a clean wheel-install MCP smoke; this is a compatibility contract, not a package-cache workaround.

An MCP server for the Retrieval academic-paper API - semantic paper search, document matching, ACE journal memory, and index inventory. Self-contained: it talks to the backend over HTTP only (just mcp + httpx), so it installs anywhere with uvx / pip - no repo checkout, no GPU, no models.

By default it targets the compute box on the lab LAN (http://10.100.100.111:8000), which trusts LAN callers so no key is needed. Off-LAN, point RETRIEVAL_API_URL at the public gateway (https://retrieval.rnarket.com) and set RETRIEVAL_API_KEY (sk-...).

Tools

Every tool's full docstring (purpose + each argument with its default + an example) is what your LLM sees - call them by name. Summary:

Paper retrieval

Journal work-memory (scoped to the current project by default)

Code KB (source stays local - only chunks are uploaded)

Public Code (admin-managed shared source corpus)

ACE playbook (accumulated, curated lessons per project)

Code KB language coverage

index_code chunks 40+ languages structurally via tree-sitter (chonkie CodeChunker): Python, TypeScript/TSX/JS/JSX (React), Java, Kotlin (incl. Jetpack Compose .kt/.kts), Swift, Go, Rust, C/C++, C#, Ruby, PHP, Lua, Scala, Dart, R, Julia, Elixir, Erlang, Haskell, OCaml, SQL, GraphQL, Protobuf, HTML, CSS/SCSS (Tailwind = CSS classes), Vue, Svelte, shell, PowerShell, Dockerfile, Terraform/HCL, CMake, YAML/JSON/TOML/XML, and more. Grammarless config/text files fall back to line-window chunks; docs (.md) and binaries are skipped (docs belong in the journal via journal_index_dir).

Singleton behavior and command-mode MCP configuration

Use the same command-style stdio entry for Claude, Codex, Grok, and other MCP clients. The normal no-argument invocation automatically starts or reuses one per-user daemon and transparently bridges just this client's JSONL session. No client URL, transport option, or manual start step is required.

Install the published package once and use the same command-style configuration for every client. Existing version-pinned uvx configurations should pin retrieval-mcp==0.4.11.

{
  "mcpServers": {
    "retrieval": {
      "command": "retrieval-mcp",
      "args": []
    }
  }
}

Existing uvx command configurations remain compatible. A once-installed retrieval-mcp avoids a package-runner wrapper per client and is the recommended singleton setup. status and stop are explicit local control commands; retrieval-mcp stdio-server is only a direct debug/compatibility server and is not the normal client configuration.

Automatic code-index refresh

index_code(...) returns a local job ID before repository walking, hashing, AST chunking, upload, GPU embedding, or Qdrant upsert completes. Poll that same ID with index_code_status() through the preparing, backend queue, and terminal phases. A second index request for the same scope reuses the active job instead of starting another scan.

With auto_refresh=True, the client then keeps one filesystem watcher for that absolute repository path. Ordinary file events hash and chunk only the touched paths; ignore-rule changes trigger a full reconcile. search_code() never waits for the watcher or indexing: it returns the last completed snapshot and reports freshness separately.

Git repositories continue to honor Git's ignore rules by default. Any directory, including non-Git projects, can add scope-relative patterns to .retrievalignore; callers can add temporary patterns with exclude_globs=["generated/**", "private.py"]. Explicit delete_code() cancels/fences stale refresh work and removes vectors, inventory, and the saved refresh policy.

Windows drive and UNC scopes use pathlib's Windows path semantics and one case-insensitive canonical identity, while POSIX scope casing remains significant. Historical mixed-case Windows scopes are included through bounded catalog aliases. path_glob is always scope-relative; both slash styles are accepted, and pathspec supplies basename, directory, **, and exact-file matching semantics. Windows extended-length/device spellings (\\?\, \\.\) are rejected with guidance to use the equivalent ordinary drive or UNC path, preventing duplicate scope identities.

Config (env)

Paper artifact transport

• get_paper_markdown(..., save_path=None) returns the complete JSON-backed Markdown to the MCP caller. With save_path, it instead uses the raw GET /api/papers/{paper_id}/markdown file route, so large Markdown is not buffered as JSON by the gateway or MCP process.
• Markdown files and PDFs use HTTP Range with validated Content-Range. If the connection closes after writing a valid prefix, the next request resumes from the exact local byte offset. Completed files are published atomically and are not overwritten unless overwrite=True.
• download_papers accepts at most 10 paper IDs, preflights all destination names, disambiguates sanitized filename collisions with a stable digest, and reuses the single-artifact Range paths with at most two concurrent transfers. Errors are reported per paper/artifact.
• The separate authenticated HTTP POST /api/papers/download-batch endpoint serves a one-shot ZIP for browser/API clients. It is capped at 10 papers and 1 GiB of source artifacts, allows at most two concurrent archive builds, and intentionally rejects Range because each generated archive is a new file.

License

MIT