wende/cicada

> [!WARNING]
> ⚠️ **Disclaimer** — Today's AI agents operate completely differently from when this project was conceived. In current benchmarks, using a code intelligence tool yields little to no advantage for frontier models. **CICADA is discontinued**, and I'd advise against using it or any other "code intelligence" tool.
---
**C**ode **I**ntelligence: **C**ontextual **A**nalysis, **D**iscovery, and **A**ttribution
**Context compaction for AI code assistants** – Give your AI structured, token-efficient access to 17+ languages including Elixir, Python, TypeScript, JavaScript, Rust, and more.
> **Up to 50% less waiting · Up to 70% less tokens · Up to 99% less explanations to do**
> **Tighter context = Better Quality**
Quick Install · Security · Developers · AI Assistants · Docs
---
Why CICADA?
**The core problem:** AI code assistants waste context on blind searches. Grep dumps entire files when you only need a function signature, leaving less room for actual reasoning.
The Context Compaction Approach
Instead of raw text dumps, CICADA gives your AI **structured, pre-indexed knowledge**:
| Traditional Search | CICADA |
|-------------------|--------|
| Grep dumps entire files | Returns only signatures + call sites |
| Misses aliased imports | Tracks all reference types |
| No semantic understanding | Keyword search finds verify_credentials when you ask for "authentication" |
What You Get
- **AST-level indexing** – Module/function/class definitions with signatures, specs, docs
- **17+ language support** – Elixir, Python, TypeScript, JavaScript, Rust, Go, Java, Kotlin, Scala, C/C++, Ruby, C#, Visual Basic, Dart, PHP, Erlang (beta)
- **Complete call-site tracking** – Aliases, imports, dynamic references across all supported languages
- **Semantic search** – Find code by concept with keyword extraction or embeddings (Ollama integration)
- **Git + PR attribution** – Surface *why* code exists, not just what
- **Dependency analysis** – Bidirectional tracking (what calls this, what does this call)
- **Automatic language detection** – Works seamlessly across polyglot codebases
---
Install
Try before installing permanently
Runs CICADA on demand (worse indexing quality, but zero install).
or
Uses your editor's built-in MCP management to install CICADA.
**Available commands after installation:**
- cicada [claude|cursor|vs|gemini|codex|opencode|zed] - One-command interactive setup per project
- cicada-mcp - MCP server (auto-started by editor)
- cicada serve - Start REST API server for HTTP access to all MCP tools
- cicada status - Show index status, PR index, link status, agent files, MCP configs
- cicada stats [repo] - Display usage statistics (tool calls, tokens, execution times)
- cicada watch - Watch for file changes and automatically reindex
- cicada index - Re-index code with custom options (-f/--force, --keywords, --embeddings, --watch)
- cicada index-pr - Index pull requests for PR attribution
- cicada run [tool] - Execute any of the 7 MCP tools directly from CLI
- cicada agents install - Install Claude Code agents to ./.claude/ directory
- cicada link [parent_dir] - Links current repository to an existing index
- cicada clean - Completely removes cicada integration from your folder as well as all settings
Ask your assistant:
---
Privacy & Security
- **100% local:** parsing + indexing happen on your machine; no external access.
- **No telemetry:** CICADA doesn't collect usage or any telemetry.
- **Read-only tools:** MCP endpoints only read the index; they can't change your repo.
- **Optional GitHub access:** PR features rely on gh and your existing OAuth token.
- **Data layout:**
Your repo only gains an editor config (.mcp.json, .cursor/mcp.json, .vscode/settings.json, .gemini/settings.json, .codex/mcp.json, or .opencode.json).
---
For Developers
> Wire CICADA into your editor once, and every assistant session inherits the context.
Install & Configure