CodeWeaver gives Claude and other AI agents precise context from your codebase. Not keyword grep. Not whole-file dumps. Actual structural understanding through hybrid semantic search.
You, or Claude, or your intern, can ask questions like:
- "Where do we handle OAuth tokens?"
- "Find all API endpoint definitions"
- "Show me error handling in the payment flow"
CodeWeaver returns the exact functions, classes, and code blocks — even in unfamiliar languages or massive repositories.
Example:
Without CodeWeaver:
Claude: "Let me search for 'auth'... here are 50 files mentioning authentication"
Result: Generic code, wrong context, wasted tokens
With CodeWeaver:
You: "Where do we validate OAuth tokens?"
Claude gets: The exact 3 functions across 2 files, with surrounding context
Result: Precise answers, focused context, actual understanding
⚠️ Alpha Release: This works, but it's early. Use it, break it, help shape it.
| Feature | CodeWeaver | Serena | Cursor | Copilot Workspace | Sourcegraph Cody | Continue.dev | Bloop | Aider |
|---|---|---|---|---|---|---|---|---|
| Approach | Semantic search | Symbol lookup (LSP) | Semantic | Semantic | Keyword | Semantic | Semantic | Repo maps |
| Tool Count | 1 | 20+ | N/A | N/A | N/A | N/A | N/A | N/A |
| Prompt Overhead | ~500 tokens | ~16,000 tokens | N/A | N/A | N/A | N/A | N/A | N/A |
| Search Speed | Moderate (embeddings) | Very fast (LSP) | Moderate | Server-side | Fast | Moderate | Fast | On-demand |
| Embedding Providers | 17 | 0 (no embeddings) | 1-2 | 1 | 0 (deprecated) | 4-5 | 1 | 0 |
| Language Support | 166+ | ~30 (LSP required) | ~50-100 | All (text) | All | ~165 | Unknown | ~165+ |
| Requires Language Server | ❌ No | ✅ Yes | ❌ No | ❌ No | ❌ No | ❌ No | ❌ No | ❌ No |
| Symbol Precision | ✅ Exact symbols | ✅ Exact | ||||||
| Concept Search | ✅ Yes | ❌ Symbols only | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes | ❌ No | |
| Editing Capabilities | ❌ No | ✅ Yes (9 tools) | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes | ❌ No | ✅ Yes |
Notes:
- Serena tool count: Varies by context (20+ in claude-code, up to 35 total available)
- Serena prompt overhead: Measured with 21 active tools in claude-code context (~16,000 tokens)
- Language counts: CodeWeaver supports 166+ unique languages (27 with AST parsing, 139 with intelligent delimiter-based chunking)
📊 See detailed competitive analysis →
# Add CodeWeaver to your project
uv add --prerelease allow --dev code-weaver
# Initialize config and MCP setup
cw init
# Verify setup
cw doctor
# Start the server
cw server📝 Note:
cw initdefaults to CodeWeaver'srecommendedprofile:
- 🔑 Voyage AI API key (generous free tier)
- 🗄️ Qdrant instance (cloud or local, both free options)
Want full offline? Use
cw init --profile quickstartfor local-only operation.
🐳 Prefer Docker? See Docker setup guide →
To watch and handle your files, CodeWeaver always runs an HTTP server. You can connect to that or use your typical stdio setup:
cw init adds CodeWeaver to your project's .mcp.json:
{
"mcpServers": {
"codeweaver": {
"type": "stdio",
"cmd": "uv",
"args": ["run", "codeweaver", "server"],
"env": {"VOYAGE_API_KEY": "your-key-here"}
}
}
}or with http:
{
"mcpServers": {
"codeweaver": {
"type": "http",
"url": "http://127.0.0.1:9328"
}
}
}
|
|
|
|
|
|
I started building CodeWeaver because I believe AI agents need better context infrastructure. Right now:
- Agents re-read the same huge files repeatedly
- They get shallow, text-based context instead of structural understanding
- They are mostly given tools built for humans, not for how they actually work
- You don't control what context they see or how they get it
CodeWeaver addresses this with one focused capability: structural + semantic code understanding that you control and can deploy however you want.
Is this solving a big problem? We think so. But we're in alpha; we're probably not there yet. We also need real-world usage to prove it. That's where you come in. Use it, make it better. Worst case -- it's a good tool, best case -- you get better results and cut costs on AI.
📖 Read the detailed rationale →
Built with ❤️ by Knitli