context-server 2026.716.1

Lightweight MCP server for semantic search over organizational markdown

context-server

A lightweight MCP server for semantic search over markdown knowledge bases.

One Rust binary. ONNX Runtime is statically linked (via ort / fastembed) — no separate libonnxruntime.so to ship. SQLite is bundled. Built for AI coding agents (Claude Code, Cursor, etc.).

Features

• Index markdown into a local SQLite vector database
• Chunk by ## / ### headings (hierarchy kept in each chunk); oversized sections are split with overlap
• Hybrid search: dense embeddings (All-MiniLM-L6-v2) + BM25, fused with reciprocal rank fusion
• MCP tools: semantic_search, list_documents, answer_question
• CLI for index / search / embed smoke tests

Input contract: feed searchable prose (markdown). Structured data (YAML, etc.) should be converted to markdown before indexing — raw YAML in code fences searches poorly.

Requirements

• Rust 1.75+ (edition 2021)
• Linux x86_64 (primary target today)
• At build/link time: a C++ standard library (libstdc++) and OpenSSL development headers if your platform needs them for native-tls

On Fedora/RHEL, if the linker cannot find -lstdc++ (only libstdc++.so.6 is installed):

mkdir -p .linker && ln -sfn /usr/lib64/libstdc++.so.6 .linker/libstdc++.so
export RUSTFLAGS="-L native=$(pwd)/.linker"

Install

pip install context-server

Platform wheels: Linux x86_64/aarch64 (manylinux_2_39, glibc 2.39+ / Ubuntu 24.04+) and macOS Apple Silicon.

Build

cargo build --release

The first embedding run downloads the MiniLM model into the local Hugging Face / fastembed cache (~tens of MB, once).

Linux wheels (Podman)

Same Containerfile CI uses (Ubuntu 24.04 / glibc 2.39 — required by current ORT prebuilts):

./scripts/build-wheel.sh   # writes dist/*.whl
VERSION=2026.716.1 ./scripts/build-wheel.sh  # optional CalVer override

Releasing (CalVer)

Versions are CalVer YYYY.MMDD.N (e.g. 2026.716.1) so they are valid for both Cargo SemVer and PyPI. pyproject.toml takes the version from Cargo.toml; release CI rewrites that from the git tag before building.

tag="$(./scripts/next-calver.sh)"
git tag -a "$tag" -m "$tag"
git push origin "$tag"    # triggers Release workflow → PyPI

setuptools-scm is not used (maturin cannot consume it).

Usage

# Preview how documents will be chunked
./target/release/context-server index --input ./docs --dry-run

# Embed and write the database
./target/release/context-server index --input ./docs --db context.db

# CLI search (hybrid by default; also --mode dense|lexical)
./target/release/context-server search --db context.db "how do we handle backports"

# MCP stdio server
./target/release/context-server serve --db context.db

Remote database (GCS)

serve and search accept a remote --db and download it into the local cache ($XDG_CACHE_HOME/context-server/dbs/..., or ~/.cache/...) before opening:

# Short form (globally unique bucket)
context-server serve --db 'gs://vme-cnv-context/latest/cnv.db'

# Project-qualified Google resource name
context-server serve --db \
  'projects/itpc-gcp-hcm-pe-eng-claude/buckets/vme-cnv-context/objects/latest/cnv.db'

Uses Application Default Credentials (gcloud auth application-default login, or GOOGLE_APPLICATION_CREDENTIALS). When a sibling {object}.sha256 exists (sha256sum format), the download is skipped if the local cache already matches; otherwise the DB is re-fetched and verified. index still writes a local path only.

Claude Code

claude mcp add --transport stdio --scope user context-server \
  -- /absolute/path/to/context-server serve --db /absolute/path/to/context.db

Re-index when content changes, then restart the MCP session so serve reloads the DB into memory. For a GCS-backed DB, point --db at the gs:// or projects/.../objects/... URI instead.

If Claude rarely calls the tools (tool search defers MCP tools), add "alwaysLoad": true to the server entry in your Claude MCP config so these tools stay visible every turn.

MCP tools

Tool	Description
semantic_search	Ranked passages with similarity scores
list_documents	Indexed chunk listing
answer_question	Top passage for a question (retrieval only, no generative QA)

Architecture

Piece	Choice
Embeddings	fastembed → All-MiniLM-L6-v2, L2-normalized (model id stored in DB)
Inference	ort (static ONNX Runtime)
Storage	rusqlite (bundled SQLite), float32 blobs
Search	Hybrid: cosine dense + BM25 → reciprocal rank fusion
Chunking	## / ### + split when text exceeds ~900 chars
MCP	rmcp stdio

See PLAN.md for design notes and roadmap.

Development

cargo test
cargo build --release

License

MIT — see LICENSE.