接入你的智能体
發布者自己的指令。不是 PlugMCP 端點。
用我們的金鑰與日誌託管一份副本。
npx -y @muvon/[email protected]Claude Code 沒有本機指令的安裝連結。複製下方指令並在終端機中執行。
claude mcp add io.github.muvon-octocode -- npx -y @muvon/octocode@0.29.2在終端機中執行。
{ "mcpServers": { "io.github.muvon-octocode": { "command": "npx", "args": [ "-y", "@muvon/[email protected]" ] } }}claude_desktop_config.json
遠端 URL 還是本機指令？ 兩種安裝方式的差異
Give your AI assistant a brain for your codebase. Octocode transforms your project into a navigable knowledge graph that Claude, Cursor, and other AI agents can search, understand, and navigate.
The Problem: AI assistants are blind to your codebase. They can't search your files, understand dependencies, or remember context across sessions.
The Solution: Octocode's MCP server gives AI agents:
Works with: Claude Desktop • Cursor • Windsurf • Any MCP-compatible AI
Now your AI assistant can:
You: "Where is authentication handled?"
AI: *searches your codebase* "Authentication is in src/middleware/auth.rs,
which imports jwt.rs for token validation and calls user_store.rs for lookup."
You: "What files depend on the payment module?"
AI: *queries knowledge graph* "src/api/handlers/payment.rs imports payment/mod.rs,
which is also used by src/workers/refund.rs and src/cron/billing.rs"
You: "Find every call site of this function"
AI: *uses LSP find-references* "process_payment() is called from 4 places:
checkout.rs:87, refund.rs:134, billing.rs:56, and tests/payment_test.rs:23"
Standard RAG treats your code as flat text chunks. It finds similar-sounding snippets but has no idea that auth_middleware.rs imports jwt.rs, calls user_store.rs, and is wired into router.rs. Octocode understands structure.
# Semantic search finds the right code
octocode search "authentication middleware"
→ src/middleware/auth.rs — Similarity: 0.9234
# The GraphRAG CLI queries the optional persisted graph
octocode config --graphrag-enabled true
octocode index
octocode graphrag get-relationships --node-id src/middleware/auth.rs
Outgoing:
imports → jwt (src/auth/jwt.rs): token validation logic
calls → user_store (src/db/user_store.rs): user lookup by token
Incoming:
imports ← router (src/router.rs): wires auth into the request pipeline
Octocode uses tree-sitter AST parsing to build a live graph of files, symbols, imports, calls, inheritance, and implementations. The MCP graphrag tool builds this graph lazily from the current source tree, without an index, embeddings, or an LLM. Optional indexed GraphRAG adds semantic file discovery, descriptions, and broader architectural relationships.
Current Source → Tree-sitter AST → Live Symbol Graph ──────────────→ MCP `graphrag`
↑ ↑
Indexed Code → Embeddings + Optional LLM → Persisted File Enrichment ─────┘
contains, imports, calls, extends, and implements relationships directly from current source[graphrag].enabled = falsesemantic_search, view_signatures, graphrag, and structural_search to any MCP-compatible clientDoc tools give AI the manual for libraries you use. Octocode gives AI the blueprint of how you put them together.
Built with Rust for performance. Local-first for privacy. Open source (Apache 2.0) for transparency.
Octocode ships a reproducible retrieval benchmark (benchmark/): 127 curated code-search queries with line-range ground truth, run against octocode's own source (pinned at b1771ba so annotations never drift). The numbers below use a fully local, no-API-key stack — jina-embeddings-v2-base-code via fastembed, no reranker — so they are a floor, not a ceiling:
Tilting RRF fusion toward the BM25/keyword signal — which carries disproportionate weight for code's exact identifiers — lifts Hit@5 by +22% and Recall@10 by +20% at zero added cost.
The benchmark also flags what doesn't help here (full 6-variant matrix in benchmark/RESULTS.md): a generic local cross-encoder reranker (bge-reranker-base) actually regressed results (Hit@5 0.732 → 0.598) — code retrieval needs a code-aware reranker (e.g. voyage:rerank-2.5), not an off-the-shelf one.
git clone https://github.com/Muvon/octocode && cd octocode
git worktree add /tmp/corpus b1771ba # pin the corpus to the ground-truth commit
CORPUS=/tmp/corpus python3 benchmark/run_matrix.py # set OCTO_BIN to use a custom binary
See benchmark/README.md for methodology and metric definitions.
# Universal installer (Linux, macOS, Windows)
curl -fsSL https://raw.githubusercontent.com/Muvon/octocode/master/install.sh | sh
# macOS with Homebrew
brew install muvon/tap/octocode
# From crates.io
cargo install octocode
# Or from source (latest)
cargo install --git https://github.com/Muvon/octocode
# Download binary from releases
# https://github.com/Muvon/octocode/releases
See Installation Guide for platform-specific instructions.
# Optional: embedding provider (defaults to local FastEmbed — no key needed)
export VOYAGE_API_KEY="your-voyage-api-key"
# Optional: LLM for commit messages, code review
export OPENROUTER_API_KEY="your-openrouter-api-key"
Get your Voyage API key: voyageai.com (free tier available)
Octocode supports multiple embedding providers:
# OpenAI
export OPENAI_API_KEY="your-key"
octocode config --code-embedding-model "openai:text-embedding-3-small"
# Jina AI
export JINA_API_KEY="your-key"
octocode config --code-embedding-model "jina:jina-embeddings-v3"
# Google
export GOOGLE_API_KEY="your-key"
octocode config --code-embedding-model "google:text-embedding-005"
See API Keys guide for all supported providers.
cd /your/project
octocode index
# → ✓ Indexing complete! 342 of 342 files processed (342 new, 0 unchanged)
# Natural language search
octocode search "authentication middleware"
# Multi-query for broader results
octocode search "auth" "middleware" "session"
# Filter by language
octocode search "database connection pool" --language rust
# Search commit history
octocode search "authentication refactor" --mode commits
Add to your MCP client config (Claude Desktop, Cursor, Windsurf):
{
"mcpServers": {
"octocode": {
"command": "octocode",
"args": ["mcp", "--path", "/your/project"]
}
}
}
Done! Your AI assistant now understands your codebase structure.
Octocode includes a built-in MCP server that exposes your codebase as tools to AI assistants. This is the primary way to use Octocode — give your AI assistant direct access to search and navigate your code.
Enable the LSP tools by starting the server with your language server:
octocode mcp --path /your/project --with-lsp="rust-analyzer"
Once connected, your AI assistant can answer questions about your codebase:
You: "Where is user authentication implemented?"
AI: *uses semantic_search* "Found in src/auth/login.rs. The authenticate() function
validates credentials against the database, generates a JWT token, and stores
the session in Redis."
You: "What files depend on the payment module?"
AI: *uses graphrag* "src/api/handlers/payment.rs imports payment/mod.rs, which is also
used by src/workers/refund.rs and src/cron/billing.rs. The payment module exports
process_payment() and validate_transaction() functions."
You: "Show me all error handling in the API layer"
AI: *uses structural_search* "Found 23 error handling patterns in src/api/:
- 15 use Result<T, ApiError> with explicit error types
- 8 use .unwrap() (potential panics in handlers/user.rs:42, handlers/auth.rs:87)
- 3 use .expect() with custom messages"
Octomind (Recommended) — Zero setup, Octocode pre-configured:
curl -fsSL https://raw.githubusercontent.com/muvon/octomind/master/install.sh | bash
octomind run developer:rust
Claude Code (CLI) — Command-line setup:
claude mcp add octocode -- octocode mcp --path /path/to/your/project
Claude Desktop / Cursor / Windsurf — Add to config:
{
"mcpServers": {
"octocode": {
"command": "octocode",
"args": ["mcp", "--path", "/path/to/your/project"]
}
}
}
Config locations:
~/Library/Application Support/Claude/claude_desktop_config.json (macOS)~/.cursor/mcp.json or Settings → MCP Servers📖 Complete MCP Client Setup Guide — Detailed instructions for 15+ clients including VS Code (Cline/Continue), Zed, Replit, and more.
17 languages with full tree-sitter AST parsing:
We measure semantic search quality using a hand-annotated ground truth dataset of 254 queries (127 code + 127 docs) with precise line-range annotations. Each query has 1–3 expected results scored by relevance.
These numbers use the full cloud stack — contextual retrieval, Voyage reranker, RaBitQ quantization — on commit b1771ba with benchmark config. For the fully local baseline and the complete variant matrix, see Retrieval Quality above and benchmark/RESULTS.md.
Missed queries (6 of 127):
Missed queries (1 of 127):
Metrics: Hit@k (did the answer appear?), MRR (how high?), NDCG@10 (are best results ranked first?), Recall@k (how many found?). See benchmark/ for methodology, scoring script, and the full dataset.
Apache License 2.0 — See LICENSE for details.
Built with 🦀 Rust by Muvon in Hong Kong
AI-powered code indexer with semantic search and knowledge graphs
{ "mcpServers": { "io.github.muvon-octocode": { "command": "npx", "args": [ "-y", "@muvon/[email protected]" ] } }}工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式，而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Octocode 透過 stdio 在本機執行。
上次檢查時間：2026-11-13。 AI-powered code indexer with semantic search and knowledge graphs