codeseek

Rust-powered code intelligence CLI for AI coding agents. Builds call graphs and hybrid semantic search indexes (Dense + Sparse + RRF + Reranker) across 7 languages. Ships as native MCP tools for Claude Code and Codex CLI.

Automatically checked · GitHub checked 2026-11-14

codeseek

Overview

CodeSeek

Code intelligence CLI tool for Claude Code. AST-based call graph analysis + semantic search — right from your terminal.

Quick Start

# Install via npm (handles setup wizard + binary download automatically)
npm install -g codeseek

# First run — interactive setup wizard configures your embedding model
codeseek

# Index your project
codeseek init

# Search code by symbol name
codeseek search main --limit 10

# Query call graph
codeseek callers main
codeseek callees process_data

# Register with Claude Code / Codex as MCP tools
codeseek install

# Check status
codeseek status

# Auto-index on git commits
codeseek install-hooks

Natural Language Code Search example

╰─$ codeseek search 'how the code embedding work'
1. get_embedding (0.7973)
   /home/do/ssd/iohub/dev/codeseek/rust-core/src/services/embedding_service.rs:0
2. EmbeddingService (0.2855)
   /home/do/ssd/iohub/dev/codeseek/rust-core/src/services/embedding_service.rs:0
3. EmbeddingData (0.1449)
   /home/do/ssd/iohub/dev/codeseek/rust-core/src/services/embedding_service.rs:0
4. EmbeddingResponse (0.1304)
   /home/do/ssd/iohub/dev/codeseek/rust-core/src/services/embedding_service.rs:0
5. default_model (0.0450)
   /home/do/ssd/iohub/dev/codeseek/rust-core/src/config.rs:0

Function Call Graph example

╰─$ codeseek callgraph apply_rerank
Call graph for 'apply_rerank' (depth=1):

== Callers (upstream, depth=1) ==
  search (/home/do/ssd/iohub/dev/codeseek/rust-core/src/services/hybrid_search.rs:210)

== Callees (downstream, depth=1) ==
  rerank (/home/do/ssd/iohub/dev/codeseek/rust-core/src/services/reranker_service.rs:331)
  config (/home/do/ssd/iohub/dev/codeseek/rust-core/src/services/hybrid_search.rs:325)

Install

npm

npm install -g codeseek

The npm package ships a lightweight JS wrapper that handles:

StepDescription
First-run wizardInteractive CLI prompts for embedding API token, model, and base URL
Binary downloadAutomatically pulls the correct Rust binary for your platform from GitHub Releases
Pass-throughAll commands (init, search, callers, etc.) are forwarded to the native binary

Supported platforms:

PlatformArchitecture
macOSarm64 (Apple Silicon), x64 (Intel)
Linuxx64

Homebrew

brew tap CodeBendKit/codeseek [email protected]:CodeBendKit/codeseek.git
brew install codeseek

From source

# install protoc
# macos: brew install protobuf
# ubuntu: sudo apt install protoc

git clone https://github.com/CodeBendKit/codeseek.git
cd codeseek
./build.sh --release

build.sh compiles both the TypeScript wrapper (dist/) and the Rust binary, then installs to ~/.codeseek/bin/.

Commands

CommandDescription
codeseekFirst-time setup wizard (configures embedding model interactively)
codeseek initBuild/update code index (full on first run, MD5-incremental thereafter)
codeseek statusIndex statistics: functions, files, last update
codeseek search <query>Symbol name search (falls back from vector → graph name match)
codeseek callers <symbol>Find functions that call this symbol
codeseek callees <symbol>Find functions this symbol calls
codeseek callgraph <symbol>Query call graph with configurable depth (bi-directional)
codeseek listList all indexed projects with paths
codeseek installRegister codeseek as MCP tools in Claude Code / Codex
codeseek uninstallRemove MCP integration
codeseek uninitDelete the current project index
codeseek install-hooksInstall git hooks (post-commit/post-merge → codeseek init)
codeseek serve --mcpStart MCP server (stdio JSON-RPC, used by Claude Code internally)

All query commands support --json for machine-readable output.

Claude Code / Codex Integration

codeseek install

Writes MCP server config to:

AgentConfig file
Claude Code~/.claude.json (global, all projects) or ./.mcp.json (project-local)
Codex CLI~/.codex/config.toml

Claude Code auto-discovers these tools after restart:

ToolCapability
codeseek_searchFind symbols by name
codeseek_callersTrace upstream callers
codeseek_calleesTrace downstream callees
codeseek_callgraphQuery call graph with configurable depth (bi-directional)
codeseek_statusCheck index health

Remove integration:

codeseek uninstall

How It Works

Index Building (codeseek init)

Source files
  → Tree-sitter AST parse (7 languages)
  → Extract functions / classes / methods
  → Batch embed via API (20 texts per call, SQLite cache)
  → Store vectors in LanceDB
  → Build BM25 index in Tantivy
  → Serialize call graph (PetCodeGraph)
  → Save to ~/.codeseek/<project_hash>/

Idempotent: first run is full build, subsequent runs compare MD5 hashes — only changed files are re-processed. Use codeseek install-hooks for automatic re-index on git commit/merge.

Hybrid Search Pipeline (codeseek search)

                        ┌─────────────────────┐
User query ────────────→│  Embedding Model    │──→ Query vector
                        └─────────────────────┘
                                  │
          ┌───────────────────────┼───────────────────────┐
          ▼                       ▼                       ▼
   ┌──────────────┐      ┌───────────────┐       ┌───────────────┐
   │ Dense Search │      │ Sparse Search │       │ Graph Search  │
   │ (LanceDB ANN)│      │ (Tantivy BM25)│       │ (PetCodeGraph)│
   └──────┬───────┘      └──────┬────────┘       └────────┬──────┘
          │                      │                        │
          └──────────────────────┼────────────────────────┘
                                 ▼
                        ┌─────────────────┐
                        │   RRF Fusion    │  ← Reciprocal Rank Fusion
                        │  (Top-20 candidates)│
                        └────────┬────────┘
                                 │
                                 ▼
                        ┌─────────────────┐
                        │    Reranker     │  ← Cross-Encoder fine re-ranking
                        │ (Qwen3-Reranker)│     scores each (query, code) pair
                        └────────┬────────┘
                                 │
                                 ▼
                        ┌─────────────────┐
                        │   Final Results  │  ← Top-5 (or Top-N)
                        └─────────────────┘
StageTechnologyRoleSpeed
Dense SearchLanceDB + Embedding ModelSemantic vector similarityFast
Sparse SearchTantivy BM25Keyword & token matchingFast
RRF FusionReciprocal Rank FusionMerge heterogeneous scores fairlyInstant
RerankerCross-Encoder (Qwen3-Reranker-4B)Full-interaction precision scoring~1-2s
FallbackPetCodeGraphGraph-based name search (no API needed)Instant

If embedding/Reranker are unavailable, the pipeline falls back gracefully to graph-based name search.

Storage

  • Config: ~/.codeseek/config.json (global, shared across all projects)
  • Index: ~/.codeseek/<md5(project_root)>/
    • project.json — Project metadata
    • graph.bin — Serialized call graph
    • embeddings.lance/ — LanceDB vector data
    • tantivy_bm25/ — BM25 full-text index
    • file_hashes.json — MD5 incremental tracking

No daemon, no HTTP server. Every command is a standalone process.

Benchmarks

End-to-end retrieval quality on public code-search benchmarks. Every number below is produced by the harnesses in evals/ — measured, not estimated.

BenchmarkWhat it measuresScaleMetricResult
MTEB(Code, v1)Natural language → function retrieval over real go / java / python source, plus CosQA101 tasks · 3,040 queriesnDCG@100.9182
RepoQARepository-level localization inside a real codebase (needle in a haystack)10 repos · 5 languages · 100 tasksaccuracy84.0%

MTEB(Code) — go 0.9448 · java 0.9101 · python 0.8994 · CosQA 0.9598, with recall@10 0.9720 and MRR@10 0.9004: the right function is almost always retrieved, so what remains to win is rank-1 ordering rather than candidate generation.

RepoQA — typescript 100% · python 90% · rust 90% · java 85% · cpp 55%.

# Full reproduction commands: evals/*/results/README.md
cd evals/mtebcode && python3 -m mtebcode_eval prepare --tasks csn,CosQA
cd evals/repoqa   && python3 -m repoqa_eval prepare

Conditions: codeseek v0.1.31 with Qwen/Qwen3-Embedding-4B, index.hybrid.cutoff.enabled = false. Full reports under evals/mtebcode/results/ and evals/repoqa/results/.

Supported Languages

LanguageFunctionsStructs/ClassesCall Graph
Rust✅✅✅
Python✅✅✅
JavaScript✅✅✅
TypeScript✅✅✅
Go✅✅✅
C/C++✅✅✅
Java✅✅✅

Configuration

~/.codeseek/config.json:

{
  "embedding": {
    "provider": "openai-compatible",
    "model": "Qwen/Qwen3-Embedding-4B",
    "api_token": "sk-...",
    "api_base_url": "https://api.siliconflow.cn/v1",
    "dimensions": 2560,
    "instruct": "Given a code search query, retrieve relevant functions or code snippets that answer the query."
  },
  "index": {
    "min_code_block_length": 16,
    "enable_reranker": true,
    "hybrid": {
      "enable_bm25": true,
      "bm25_top_k": 20,
      "vector_top_k": 20,
      "rrf_k": 60,
      "rrf_top_k": 20
    },
    "reranker": {
      "enabled": true,
      "model": "Qwen/Qwen3-Reranker-4B",
      "api_token": "sk-...",
      "api_base_url": "https://api.siliconflow.cn/v1/rerank",
      "top_n": 5,
      "candidate_multiplier": 5,
      "timeout_secs": 60
    }
  },
  "installed_hooks": {}
}

embedding.instruct — Qwen3-Embedding query-side instruction (asymmetric retrieval). Only the search query gets the official Instruct: {instruct}\nQuery:{query} prefix; indexed code stays raw, so existing vector indexes remain valid. Omit it to use the default shown above; set it to "" to turn the prefix off.

Model Roles

ModelRoleWhen
Qwen/Qwen3-Embedding-4BConverts code → vectors for dense searchIndex building
Qwen/Qwen3-Reranker-4BScores (query, code) pairs for precisionSearch time

Set via the interactive wizard on first run, or create manually.

Development

cd rust-core

# Build
cargo build

# Build + install to ~/.codeseek/bin/
cd .. && ./build.sh --release

# Run tests
cargo test

# Compile TypeScript wrapper
npm run build

License

MIT

Built with: Tree-sitter · Petgraph · LanceDB · Tantivy · Tokio · Clap