codedb
Code intelligence server for AI agents. Zig core. MCP native. Zero dependencies.
- Source repo
- justrach/codedb
- Stars
- ★ 1.4k
- Last updated
- 20d ago
- License
- BSD-3-Clause
- Primary language
- Zig
- FA score
- 60/100 · Some gaps
At a glance
- How it runs
- Works with
- Portable with changesChatGPT · Codex · Claude Code · Claude.ai · OpenAI API · Claude API
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A developer coding in Claude Code or Cursor wants to quickly locate a symbol definition or its callers; codedb provides responses in under a millisecond.
- Not a fit if
- Teams needing remote or shared deployment: localhost-only, no auth
- Users wanting built-in code editing: codedb has no edit tool
- Privacy-sensitive teams unwilling to disable telemetry, which is on by default
- Source review
- 60/100 · Some gaps
What does this agent do, and when should you use it?
codedb is a code intelligence server for AI agents, built in Zig with native MCP support. It provides structural indexing, trigram search, word index, dependency graph, and file watching capabilities. Through 20+ MCP tools and an HTTP API, it offers fast, context-rich code queries, but does not have editing capabilities, leaving editing to the client's native tools. Its unique advantage is indexing performance: pre-indexed queries can be up to 538x faster than ripgrep, and index building is extremely fast (5200 files in ~310ms cold start). It supports multiple languages including Zig, C/C++, Python, TypeScript/JavaScript, Rust, Go, and more, with a local-first privacy approach.
codedb indexes project code and provides efficient queries to AI agents. On startup, it parses the codebase, building structural outlines, trigram full-text indexes, inverted word indexes, and dependency graphs, and continuously watches for file changes to incrementally update indexes. It exposes tools via the MCP protocol (stdio) or HTTP API (localhost:7719), such as codedb_tree (file tree), codedb_outline (symbols list), codedb_search (full-text search), codedb_word (O(1) word lookup), codedb_callers (callers), codedb_deps (dependency graph), and codedb_context (task composer). Agents can call these tools to explore code without scanning the filesystem, saving time and tokens.
- A developer coding in Claude Code or Cursor wants to quickly locate a symbol definition or its callers; codedb provides responses in under a millisecond.
- An AI agent performing refactoring on a large codebase needs to understand dependencies; codedb's deps tool provides a complete graph.
- Searching for a specific string (like 'handleAuth') without full file scans; codedb's trigram search accelerates the query.
- When needing to understand project structure, codedb's tree and outline tools generate structure with symbol counts.
- For indexing multiple projects once and then querying frequently (e.g., in CI or local development), codedb's pre-indexed queries minimize latency.
- For exploring public GitHub repos remotely, use the DeepWiki integration's remote MCP tools (read_wiki_structure, ask_question).
How do you install or deploy this agent?
macOS/Linux install:
curl -fsSL https://codedb.codegraff.com/install.sh | bashThis auto-registers MCP servers in Claude Code, Codex, Gemini CLI, Cursor, Windsurf, and Devin. On Windows, run in PowerShell:
irm https://raw.githubusercontent.com/justrach/codedb/v0.2.5838/install/install.ps1 | iexAlternatively, use npm: npx -y codedeebee mcp or npm install -g codedeebee. Network access is required to download binaries.
How do you use this agent?
Start as an MCP server:
codedb mcp /path/to/your/projectOr as an HTTP server:
codedb serve /path/to/your/projectThen call tools via MCP client or HTTP, for example:
curl localhost:7719/tree
curl localhost:7719/symbol?name=AgentRegistry
curl localhost:7719/search?q=handleAuth&max=10Use codedb --version to check version, disable telemetry with CODEDB_NO_TELEMETRY=1 or --no-telemetry.
What are this agent's strengths and limitations?
- Extremely fast queries: pre-indexed queries are 538x faster than ripgrep with sub-millisecond latency
- Zero runtime dependencies, single binary, cross-platform support for macOS/Linux/Windows
- Built-in sensitive file filtering (.env, credentials, keys) and local-first data privacy
- Auto-registers MCP servers for mainstream AI tools (Claude Code, Codex, Cursor, etc.)
- Portable snapshot for instant MCP server startup
- Alpha status: API may change, snapshot format may not be compatible
- Binds to localhost only, no authentication, not suitable for remote deployment
- Some languages have only lightweight outline support, not full parsing
- No editing capability, requires integration with client's native edit tools
- Telemetry is on by default and must be explicitly disabled
How does this agent compare with similar options?
Compared to ast-grep, ripgrep, and grep, codedb is significantly faster in pre-indexed scenarios and offers unique features like structural parsing, dependency graph, and MCP protocol. Compared to ctags, codedb provides more indexing features.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| codedb This agent | 60 · Some gaps | MCP serverFree | ★ 1.4k | 20d ago | Zig | ChatGPT · Codex · Claude Code · Claude.ai · OpenAI API · Claude API |
| Empryo | 50 · Major gaps | CLIFree + model costs | ★ 1.4k | 3d ago | TypeScript | OpenAI API · Claude API |
| CodeSeek | 48 · Major gaps | CLIFree + model costs | ★ 768 | 2mo ago | Rust | Codex · Claude Code |
| Devo Desktop Coding Agent | 58 · Major gaps | Desktop appFree + model costs | ★ 338 | 3d ago | Rust | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence: toolset is read-only, no edit capability, aligns with least privilege; installer auto-registers MCP and writes configs without explicit user confirmation; data flow transparent, telemetry on by default with aggregate data; sensitive files auto-excluded; zero dependencies, but npm package downloads binary with SHA256 verification; external effects include auto-registration and telemetry with opt-out; rollback and snapshots supported; source attribution clear but publisher unverified. Deductions: insufficient user confirmation, telemetry default-on.
Evidence: README consistent with code structure, architecture clear; zero dependencies, but npm package relies on network; failure messages not detailed. Deductions: inadequate failure messages.
Evidence: targets AI agents, scenarios clear; capability boundaries clear, no edit; triggers precise, tool names clear; cross-platform support. Deductions: none significant.
Evidence: good documentation structure, detailed install notes; naming stable but API may change; examples abundant; known limitations clear; BSD-3-Clause license; versioning and changelog incomplete; maintenance responsibility clear. Deductions: naming stability insufficient, versioning/changelog incomplete.
Evidence: structured output saves tokens; high marginal value with unique features; good cost-benefit with zero dependencies. Deductions: none significant.
Evidence: performance claims have benchmark data but no reproduction details; limited cross-source verification; facts and inferences not clearly separated. Deductions: insufficient verifiability.
- Telemetry is on by default; users must opt out.
- Installer auto-modifies multiple tool configs; review required.
- Publisher identity unverified; use with caution.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.2.5860 CodeDB v0.2.5860
Release notes excerpt
Faster hosted semantic index builds codedb semantic-index now sends 100-item batches with eight requests in flight by default. CODEDB_SEMANTIC_INDEX_CONCURRENCY still overrides the concurrency (1-8). Index builds wait out hosted rate limits with a doubling backoff capped at 30 seconds instead of failing after a few quick retries. Interactive codedb context calls are unchanged and still fall back to local results immediately. Local retrieval, the watcher, and the installer are unchanged.View on GitHub ↗ -
v0.2.5859 CodeDB v0.2.5859
Release notes excerpt
Hosted semantic retrieval Fixes codedb context reporting semantic lane: unavailable with the default hosted configuration. The hosted default model id is now codedb-hosted-v1. Local ANN sidecars built by earlier releases are rebuilt once in the background; the bounded exact rerank serves requests meanwhile. Custom endpoints and explicit model overrides are unchanged. Local retrieval, the watcher, and the installer are unchanged.View on GitHub ↗ -
v0.2.5857 CodeDB v0.2.5857
Release notes excerpt
More MCP clients The installer now registers codedb in every detected MCP client, not only Claude Code, Codex, Gemini CLI, Cursor, Windsurf, and Devin. New when the client is already present: oh-my-pi, Hermes, Qwen Code, ZCode, Trae, Cline, Copilot CLI, Antigravity, Kiro, and OpenCode. Malformed JSON is left untouched. codedb nuke removes the same entries. Antigravity writes only ~/.gemini/antigravity/mcp_config.json, not Gemini CLI settings. Nuke still clears the stray Gemini path and the non-CN Trae User/mcp.json path. Retrieval, the watcher, and hosted embeddings are unchanged. …View on GitHub ↗