Opencodesearch

opencodesearch is an asynchronous Rust code search system with a Model Context Protocol (MCP) server. It indexes large repositories into vector + keyword backends and serves search results through MCP tools.

Features

• Fully async runtime (tokio)
• 4 isolated processes:
  • orchestrator (state machine + supervision)
  • background ingestor
  • MCP server process
  • git watchdog process
• Required crates integrated and used in runtime code:
  • opencodesearchparser
  • qdrant-client
  • ollama-rs
  • rmcp
• Hybrid retrieval:
  • semantic search (Qdrant vectors)
  • keyword search (Quickwit HTTP + local shadow fallback)
• MCP server compatible with MCP clients using streamable HTTP and stdio transports

Architecture

State machine in orchestrator:

• SPINUP: load config.json
• NORMAL: run ingestor + mcp + watchdog
• UPDATE: keep watchdog, stop ingestor + mcp during update window
• CLOSING: stop all children gracefully

Update flow:

• watchdog tracks git commits since last sync
• when threshold (commit_threshold) is reached:
  • send UPDATE_START to orchestrator
  • pull + compute changed/deleted files
  • remove stale docs
  • reindex changed files
  • send UPDATE_END

Requirements

• Rust stable toolchain
• Docker + Docker Compose
• Local network access to:
  • Ollama (11434)
  • Qdrant (6333 HTTP, 6334 gRPC)
  • Quickwit (7280)

Configuration

config.json schema:

{
  "codebase": {
    "directory_path": "/path/to/massive/repo",
    "git_branch": "main",
    "commit_threshold": 50,
    "mcp_server_name": "My cool codebase",
    "mcp_server_url": "http://localhost:9443",
    "background_indexing_threads": 2
  },
  "ollama": {
    "server_url": "http://localhost:11434",
    "embedding_model": "qwen3-embedding:0.6b",
    "context_size": 2000
  },
  "qdrant": {
    "server_url": "http://localhost:6334",
    "collection_name": "opencodesearch-code-chunks",
    "api_key": null
  },
  "quickwit": {
    "quickwit_url": "http://localhost:7280",
    "quickwit_index_id": "opencodesearch-code-chunks"
  }
}

Important:

• qdrant.server_url should target the gRPC endpoint port (6334) for qdrant-client.
• quickwit.quickwit_url should target HTTP (7280).

Start Backend Services

Run all local dependencies:

docker compose up -d

Check containers:

docker ps

Running the System

1) Orchestrator mode (recommended)

Starts and supervises all child processes.

cargo run -- orchestrator --config config.json

2) Individual process modes

You can run each process directly for debugging.

Ingestor:

cargo run -- ingestor --config config.json

MCP server:

cargo run -- mcp --config config.json

MCP server over stdio (for local MCP clients):

cargo run -- mcp-stdio --config config.json

Watchdog (requires orchestrator IPC env):

OPENCODESEARCH_IPC_SOCKET=/tmp/opencodesearch.sock cargo run -- watchdog --config config.json

MCP Server Usage

The MCP server supports:

• streamable HTTP via cargo run -- mcp --config config.json
• stdio via cargo run -- mcp-stdio --config config.json

Implemented MCP tool:

• search_code
  • input:
    • query: string
    • limit?: number (default 8, max 50)
  • output (structured JSON): array of objects with
    • snippet
    • path
    • start_line
    • end_line
    • score
    • source

Example tool input

{
  "query": "which function changes obj variable",
  "limit": 5
}

Result shape

{
  "hits": [
    {
      "path": "/repo/module.py",
      "snippet": "def mutate(obj): ...",
      "start_line": 10,
      "end_line": 22,
      "score": 0.92,
      "source": "qdrant"
    }
  ]
}

Using With MCP Clients

This server supports both:

• streamable HTTP (cargo run -- mcp --config config.json)
• local stdio (cargo run -- mcp-stdio --config config.json)

OpenAI Codex

Codex supports both stdio and streamable HTTP MCP servers.

Stdio (CLI):

codex mcp add opencodesearch -- \
  cargo run --quiet --manifest-path /home/brosnan/opencodesearch/Cargo.toml -- \
  mcp-stdio --config /home/brosnan/opencodesearch/config.json

Remote HTTP (~/.codex/config.toml or .codex/config.toml):

[mcp_servers.opencodesearch]
url = "http://localhost:9443/"

Then verify:

codex mcp list

OpenCode

OpenCode config uses the mcp section in opencode.json (or opencode.jsonc).

Remote HTTP:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "opencodesearch": {
      "type": "remote",
      "url": "http://localhost:9443/",
      "enabled": true
    }
  }
}

Local stdio:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "opencodesearch": {
      "type": "local",
      "command": [
        "cargo",
        "run",
        "--quiet",
        "--manifest-path",
        "/home/brosnan/opencodesearch/Cargo.toml",
        "--",
        "mcp-stdio",
        "--config",
        "/home/brosnan/opencodesearch/config.json"
      ],
      "enabled": true
    }
  }
}

Claude Code

Claude Code supports HTTP, SSE, and stdio MCP transports.

Remote HTTP:

claude mcp add --transport http opencodesearch http://localhost:9443/

Local stdio:

claude mcp add --transport stdio opencodesearch -- \
  cargo run --quiet --manifest-path /home/brosnan/opencodesearch/Cargo.toml -- \
  mcp-stdio --config /home/brosnan/opencodesearch/config.json

Then verify:

claude mcp list

TLS / HTTPS Notes

• Default local config uses http://localhost:9443.
• For https://..., provide a certificate trusted by your MCP client.
• TLS cert and key defaults:
  • certs/localhost-cert.pem
  • certs/localhost-key.pem
• Override TLS file paths with:
  • OPENCODESEARCH_TLS_CERT_PATH
  • OPENCODESEARCH_TLS_KEY_PATH
• For Codex specifically, you can provide a custom CA bundle with CODEX_CA_CERTIFICATE.

References:

• Codex MCP docs: https://developers.openai.com/codex/mcp
• OpenCode MCP docs: https://opencode.ai/docs/mcp-servers/
• Claude Code MCP docs: https://code.claude.com/docs/en/mcp

Quick curl test

Use the included script:

./test_mcp_curl.sh

Optional:

• MCP_URL=https://localhost:9443/ MCP_INSECURE=1 ./test_mcp_curl.sh

The script performs the required MCP HTTP handshake steps:

1. initialize
2. extract mcp-session-id from response headers
3. send notifications/initialized with the same mcp-session-id
4. call tools/call for search_code

Manual curl sequence

Initialize and capture session id:

curl -sS -D headers.txt http://localhost:9443/ \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"curl-test","version":"1.0"}}}'

Send initialized notification:

SESSION_ID="$(awk 'tolower($1)=="mcp-session-id:"{print $2}' headers.txt | tr -d '\r' | tail -n 1)"
curl -sS http://localhost:9443/ \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -H "mcp-session-id: ${SESSION_ID}" \
  -d '{"jsonrpc":"2.0","method":"notifications/initialized"}'

Call the MCP tool:

curl -N http://localhost:9443/ \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -H "mcp-session-id: ${SESSION_ID}" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"search_code","arguments":{"query":"which function mutates obj","limit":5}}}'

Rust API Documentation

The crate exposes reusable modules for embedding, indexing, MCP serving, and process control.

Modules

• config: parse typed app config (AppConfig)
• chunking: parse/split source files into chunks (chunk_file)
• indexing: indexing runtime (IndexingRuntime)
• qdrant_store: vector storage + semantic query (QdrantStore)
• quickwit: keyword storage/query (QuickwitStore)
• mcp: MCP server type (OpenCodeSearchMcpServer)
• watchdog: git update monitor (WatchdogProcess)
• orchestrator: multi-process supervisor (Orchestrator)

Minimal Rust indexing example

use opencodesearch::config::AppConfig;
use opencodesearch::indexing::IndexingRuntime;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let config = AppConfig::from_path("config.json")?;
    let runtime = IndexingRuntime::from_config(config)?;

    runtime.index_entire_codebase().await?;
    Ok(())
}

Minimal Rust semantic search example

use opencodesearch::config::AppConfig;
use opencodesearch::indexing::IndexingRuntime;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let config = AppConfig::from_path("config.json")?;
    let runtime = IndexingRuntime::from_config(config)?;

    let query_vec = runtime.embed_query("where is object mutated") .await?;
    let hits = runtime.qdrant.semantic_search(query_vec, 5).await?;

    for hit in hits {
        println!("{}:{}-{}", hit.path, hit.start_line, hit.end_line);
    }

    Ok(())
}

Minimal Rust MCP server embedding

use opencodesearch::config::AppConfig;
use opencodesearch::indexing::IndexingRuntime;
use opencodesearch::mcp::OpenCodeSearchMcpServer;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let config = AppConfig::from_path("config.json")?;
    let runtime = IndexingRuntime::from_config(config)?;
    OpenCodeSearchMcpServer::new(runtime)
        .run_streamable_http("http://localhost:9443")
        .await
}

Testing

Standard tests

cargo test

Live container integration tests

Requires running Docker services and local git:

cargo test -- --ignored

Current ignored integration tests validate:

• Ollama connectivity
• Quickwit + Qdrant connectivity
• full indexing flow on generated Python project
• retrieval through MCP search path with non-exact query phrasing
• 100-commit refactor scenario for watchdog threshold behavior

Troubleshooting

• Quickwit health endpoint: use http://localhost:7280/health/livez
• If embeddings fail, confirm Ollama model availability:
  • qwen3-embedding:0.6b
• Qdrant client requires gRPC port (6334) in config
• If integration tests fail on startup race, rerun after a short container warmup