context-server

Semantic search over a folder of markdown, served as an MCP server for coding agents.

Index once into a SQLite DB (embeddings + BM25). Point Claude Code, Cursor, or any MCP client at serve, and the agent can search that corpus instead of guessing from memory.

One Rust binary. ONNX Runtime is linked in via ort / fastembed — no separate libonnxruntime to ship. SQLite is bundled.

Quick start

pip install context-server
# or: uvx context-server@latest …

context-server index --input ./docs --db context.db
context-server search --db context.db "how do we handle backports"
context-server serve --db context.db

Wheels: Linux x86_64/aarch64 (manylinux_2_39 / glibc 2.39+, e.g. Ubuntu 24.04+) and macOS Apple Silicon.

The first embedding run downloads BGE-small-en-v1.5 into $XDG_CACHE_HOME/context-server/fastembed/ (or ~/.cache/...; once, tens of MB). Override with FASTEMBED_CACHE_DIR or HF_HOME.

Optional: tell the agent when to use this corpus

context-server index --input ./docs --db context.db \
  --instructions-file ./mcp-instructions.txt
# or: --instructions 'Use semantic_search for questions about …'

That text is stored in the DB and exposed as MCP ServerInfo.instructions when you serve.

Claude Code

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

--refresh + @latest rechecks PyPI on each start. If Claude rarely surfaces the tools, set "alwaysLoad": true on the server entry in your Claude MCP config.

Cursor

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

{
  "mcpServers": {
    "context-server": {
      "command": "uvx",
      "args": [
        "--refresh",
        "context-server@latest",
        "serve",
        "--db",
        "/absolute/path/to/context.db"
      ]
    }
  }
}

Reload MCP after editing. Re-index when content changes, then restart the MCP session so serve reloads the DB.

What it indexes

Only .md / .markdown. Chunks on # / ## / ###, keeps the heading path on each chunk, and splits long sections with overlap.

Convert structured sources (YAML, etc.) to prose before indexing. Fenced YAML searches poorly; a short paragraph that keeps names, roles, and relationships together works much better.

Try the sample set:

cargo build --release
./target/release/context-server index --input examples/sample-docs --dry-run
./target/release/context-server index --input examples/sample-docs --db /tmp/sample.db
./target/release/context-server search --db /tmp/sample.db "password reset"

Search

Default mode is hybrid: dense cosine (BGE-small-en-v1.5) plus BM25, fused with reciprocal rank fusion. Dense catches paraphrase; BM25 catches exact tokens (usernames, acronyms, IDs).

context-server search --db context.db --mode hybrid "query"   # default
context-server search --db context.db --mode dense "query"
context-server search --db context.db --mode lexical "query"

# Scope to a subtree / heading / metadata tag
context-server search --db context.db --path-prefix teams/ "who owns storage"
context-server search --db context.db --heading Backport "z-stream"
context-server get --db context.db --path teams/storage.md --chunk 0

MCP tools

Search hits cite chunks as source_path#chunk_index. Call get_document to pull the full text for quoting.

Remote database (GCS)

serve and search accept a gs:// URI. The object is cached under $XDG_CACHE_HOME/context-server/dbs/ (or ~/.cache/...). index still writes a local path only.

context-server serve --db 'gs://my-bucket/latest/context.db'

# Project-qualified form also works (gs:// required; stripped for the Storage API)
context-server serve --db \
  'gs://projects/my-gcp-project/buckets/my-bucket/objects/latest/context.db'

Uses Application Default Credentials. If a sibling {object}.sha256 exists (sha256sum format), a matching local cache is reused; otherwise the DB is re-fetched and verified.

CLI

context-server index  --input <path> [--db FILE] [--dry-run] [--batch N]
                      [--full] [--sync]
                      [--instructions TEXT | --instructions-file FILE]
context-server serve  --db <local path | gs://…>
context-server search --db <local path | gs://…> [--limit N] [--mode hybrid|dense|lexical]
                      [--path-prefix P] [--heading H] [--tag T] <query>
context-server get    --db <local path | gs://…> --path FILE [--chunk N]
context-server embed  <query>         # smoke-test query embedding (BGE instruction)

index is upsert-only by default. Use --sync only when the database should exactly mirror the current input: it deletes indexed paths missing from that input, and an empty input removes every indexed document. The former --update behavior is now the default; replace previous prune-by-default commands with an explicit --sync.

Build from source

cargo build --release
cargo test

Rust 1.90+, Linux x86_64 is the primary target. You need a C++ stdlib for the linker (libstdc++) and whatever OpenSSL/native-tls needs on your platform.

On Fedora/RHEL, if the linker wants -lstdc++ but only libstdc++.so.6 exists:

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

Linux wheels (same image CI uses — Ubuntu 24.04 / glibc 2.39):

./scripts/build-wheel.sh
VERSION=2026.716.1 ./scripts/build-wheel.sh   # optional override

Releasing

CalVer YYYY.MMDD.N (e.g. 2026.716.1) so versions work for both Cargo and PyPI. Run the Release workflow on main (Actions UI or CLI); it picks the next version, builds wheels, publishes to PyPI, then creates the matching git tag and GitHub Release (with wheels attached).

gh workflow run release.yml --repo context-server/context-server

Design notes

Under the hood: fastembed BGE-small-en-v1.5 (384-d, L2-normalized; query instruction applied at search time), rusqlite with float32 blobs, rmcp over stdio. index is incremental by file: unchanged files (same post-chunk content hash) are skipped, so the embedding model is not loaded. Indexing safely upserts by default. Pass --sync to also remove database paths missing from --input, or --full to re-embed everything collected. A model or chunker migration requires a complete-corpus --sync run.

More detail and roadmap: PLAN.md.

Supported scale

The primary target is up to 10,000 chunks; 50,000 chunks is the regularly benchmarked upper range for the exact in-memory implementation. Re-evaluate storage/index architecture around 100,000 chunks or 500 MiB resident memory. Run scripts/benchmark-scale.py --source-db context.db to reproduce structural latency and database-size measurements.

License

MIT — see LICENSE.