CICADA MCP Server

<div align="center"> <img src="https://github.com/user-attachments/assets/699c0a3f-dfae-4e14-81db-c8ba970d9a35">

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.

---

Code Intelligence: Contextual Analysis, Discovery, and Attribution

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

</div>

---

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

# 1. Install uv (if needed)
# curl -LsSf https://astral.sh/uv/install.sh | sh
uv tool install cicada-mcp

# In your repo
cicada claude   # or: cicada cursor, cicada vs, cicada gemini, cicada codex, cicada opencode, cicada zed

<div align="left"> <summary><strong>Try before installing permanently</strong></summary> Runs CICADA on demand (worse indexing quality, but zero install).

uvx cicada-mcp claude   # or cursor, vs

or

claude mcp add cicada uvx cicada-mcp

gemini mcp add cicada uvx cicada-mcp

codex mcp add cicada uvx cicada-mcp

kimi mcp add --transport stdio cicada -- cicada-mcp

Uses your editor's built-in MCP management to install CICADA.

</details> </div>

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:

# Elixir
"Show me the functions in MyApp.User"
"Where is authenticate/2 called?"

# Python
"Show me the AuthService class methods"
"Where is login() used in the codebase?"

# Both languages
"Find code related to API authentication"

---

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:

  ~/.cicada/projects/<repo_hash>/
  ├─ index.json      # modules, functions, call sites, metadata
  ├─ config.yaml     # indexing options + mode
  ├─ hashes.json     # incremental indexing cache
  └─ pr_index.json   # optional PR metadata + reviews
  

  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

cd /path/to/project
cicada claude   # or cicada cursor / cicada vs / cicada gemini / cicada codex / cicada opencode / cicada zed

Enable PR Attribution (optional)

brew install gh    # or apt install gh
gh auth login
cicada index-pr .     # incremental
cicada index-pr . --clean   # full rebuild

Unlocks questions like "Which PR introduced line 42?" or "What did reviewers say about billing.ex?"

Automatic Re-indexing with Watch Mode

Enable automatic reindexing when files change by starting the MCP server with the --watch flag:

** .mcp.json**

{
  "mcpServers": {
    "cicada": {
      "command": "cicada-mcp",
      "args": ["--watch"],
      "env": {
        "CICADA_CONFIG_DIR": "/home/user/.cicada/projects/<hash>"
      }
    }
  }
}

When watch mode is enabled:

• A separate process monitors .ex, .exs (Elixir) and .py (Python) files for changes
• Changes are automatically reindexed (incremental, fast)
• 2-second debounce prevents excessive reindexing during rapid edits
• The watch process stops automatically when the MCP server stops
• Excluded directories: deps, _build, node_modules, .git, assets, priv, .venv, venv

CLI Cheat Sheet

Note: Language detection is automatic – CICADA detects Elixir (mix.exs) and Python (pyproject.toml) projects automatically.

Command	Purpose	Run When
cicada claude	Configure MCP + incremental re-index	First setup, after local changes
cicada status	Check index health, link status, agent files	After setup, troubleshooting
cicada stats	View usage statistics and token metrics	Monthly reviews, optimization
cicada watch	Monitor files and auto-reindex on changes	During active development
cicada index --keywords .	Rebuild with keyword indexing	After large refactors or enabling keywords mode
cicada index --embeddings .	Rebuild with embeddings (semantic search)	When you want Ollama-powered semantic analysis
cicada index-pr .	Sync PR metadata/reviews	After new PRs merge

Troubleshooting

<details> <summary><b>"Index file not found"</b></summary>

Run the indexer first:

cicada index /path/to/project

Ensure indexing completed successfully. Check for ~/.cicada/projects/<hash>/index.json.

</details> <details> <summary><b>"Module not found"</b></summary>

Use the exact module name as it appears in code (e.g., MyApp.User, not User).

If module was recently added, re-index:

cicada index .

</details> <details> <summary><b>MCP Server Won't Connect</b></summary>

Troubleshooting checklist:

1. Verify configuration file exists:

   # For Claude Code
   ls -la .mcp.json
   
   # For Cursor
   ls -la .cursor/mcp.json
   
   # For VS Code
   ls -la .vscode/settings.json
   

2. Check paths are absolute:

   cat .mcp.json
   # Should contain: /absolute/path/to/project
   # Not: ./project or ../project
   

3. Ensure index exists:

   ls -la ~/.cicada/projects/
   # Should show directory for your project
   

4. Restart editor completely (not just reload window)

5. Check editor MCP logs:

   • Claude Code: --debug
   • Cursor: Settings → MCP → View Logs
   • VS Code: Output panel → MCP

</details> <details> <summary><b>PR Features Not Working</b></summary>

Setup GitHub CLI:

# Install GitHub CLI
brew install gh  # macOS
sudo apt install gh  # Ubuntu
# or visit https://cli.github.com/

# Authenticate
gh auth login

# Index PRs
cicada index-pr

Common issues:

• "No PR index found" → Run cicada index-pr .
• "Not a GitHub repository" → Ensure repo has GitHub remote
• Slow indexing → First-time indexing fetches all PRs; subsequent runs are incremental
• Rate limiting → GitHub API has rate limits; wait and retry if you hit limits

Force rebuild:

cicada index-pr --clean

</details> <details> <summary><b>Keyword Search Not Working</b></summary>

Error: "Keyword search not available"

Cause: Index was built without keyword extraction.

Solution:

# Re-index with keyword extraction
cicada index .  # or --keywords

Verify:

cat ~/.cicada/projects/<hash>/config.yaml
# Should show:
# indexing:
#   mode: keywords

</details>

More detail: PR Indexing, Incremental Indexing.

<details> <summary><b>Python Indexing</b></summary>

Requirements:

• Node.js (for scip-python indexer)
• Python project with pyproject.toml

First-time setup: CICADA automatically installs scip-python via npm on first index. This may take a minute.

Known limitations (Beta):

• First indexing may be slower than Elixir (SCIP generation step)
• Large virtual environments (.venv) are automatically excluded
• Some dynamic Python patterns may not be captured

Performance tips:

# Ensure .venv is excluded
echo "/.venv/" >> .gitignore

# Use keywords mode for quickest indexing
cicada index --keywords .

Report issues: GitHub Issues with "Python" label

</details>

---

For AI Assistants

CICADA ships 7 focused MCP tools designed for efficient code exploration across Elixir, Python, and Erlang codebases.

🧭 Which Tool Should You Use?

Need	Tool	Notes
Start exploring	query	🚀 START HERE - Smart discovery with keywords/patterns + filters (scope, recent, path)
View a module's complete API	search_module	Functions, signatures, specs, docs. Use what_calls_it/what_it_calls for bidirectional analysis
Find where a function is used	search_function	Definition + all call sites. Supports wildcards (*) and OR (|) patterns
Track git history	git_history	Unified tool: blame, commits, PRs, function evolution (replaces 4 legacy tools)
Drill down into results	expand_result	Auto-expands modules or functions from query results
Advanced index queries	query_jq	Custom jq queries for power users

Want to see these tools in action? Check out Complete Workflow Examples with pro tips and real-world scenarios.

Core Tools

query - Smart code discovery (your starting point)

• Automatically detects keywords vs patterns
• Filters: scope (public/private), recent (last 14 days), filter_type (modules/functions), match_source (docs/strings)
• Returns snippets with smart next-step suggestions
• Use path_pattern to filter by location

search_module - Deep module analysis

• View complete API: functions, signatures, specs, docs
• For Python: Shows classes with method counts and signatures
• For Elixir: Shows functions with arity notation
• Bidirectional analysis:
  • what_calls_it=true → See who uses this module (impact analysis)
  • what_it_calls=true → See what this module depends on
• Supports wildcards (Elixir: MyApp.*, Python: api.handlers.*) and OR patterns (MyApp.User|MyApp.Post)
• Filter by visibility (public/private/all)

search_function - Function usage tracking

• Find definitions and all call sites
• what_calls_it=true (default) → See all callers
• what_it_calls=true → See all dependencies
• Include code examples with include_usage_examples=true
• Filter by usage_type: source, tests, or all

Git History (Unified Tool)

git_history - All git operations in one tool

• Single line: git_history("file.ex", start_line=42) → blame + PR
• Line range: git_history("file.ex", start_line=40, end_line=60) → grouped blame
• Function tracking: git_history("file.ex", function_name="create_user") → evolution
• File history: git_history("file.ex") → all PRs/commits
• Time filtering: recent=true (14d), recent=false (>14d), recent=null (all)
• Author filtering: author="john"
• Automatic PR index integration when available

Additional Tools

expand_result - Drill down from query results

• Auto-detects module vs function
• Shows complete details with usage examples
• Configure what to include: code, dependencies, callers
• Convenient wrapper around search_module and search_function

query_jq - Advanced index queries

• Direct jq queries against the index
• Schema discovery with | schema
• Compact (default) or pretty output
• Sample mode for large results

Detailed parameters + output formats: MCP_TOOLS_REFERENCE.md.

Token-Friendly Responses

All tools return structured Markdown/JSON snippets (signatures, call sites, PR metadata) instead of full files, keeping prompts lean.

New in v0.5.1: All tools now use compact output by default to minimize token usage. Use verbose=true for detailed output with full docs and specs.

---

---

Documentation

• Codebook – Complete feature reference and user guides
• Workflows – Real-world examples chaining tools together
• Installation – Step-by-step setup for all editors
• Contributing – Development guidelines and architecture
• CHANGELOG.md – Release notes

Deep Dives:

• Keyword Extraction Analysis – Semantic search internals
• PR Indexing – GitHub integration details
• MCP Tool Call Benchmarking – Token/time benchmarks
• Tool Discoverability – UX improvements research

---

Roadmap

Current Status

Production Ready:

• ✅ Elixir (tree-sitter)
• ✅ Python (SCIP)
• ✅ TypeScript (SCIP)
• ✅ JavaScript (SCIP)
• ✅ Rust (SCIP)

Beta:

• 🚧 Erlang (tree-sitter)
• 🚧 Go (SCIP)
• 🚧 Java/Kotlin/Scala (SCIP)
• 🚧 C/C++ (SCIP)
• 🚧 Ruby (SCIP)
• 🚧 C#/Visual Basic (SCIP)
• 🚧 Dart (SCIP)
• 🚧 PHP (SCIP)

---

Comparison to Alternatives

Feature	CICADA	Serena	Codicil (Elixir-only)
Analysis Method	SCIP (static index)	LSP (real-time server)	LLM summaries + embeddings
Code Editing	❌	✅	❌
Git Context	✅ PR history, blame, evolution	❌	❌
Resource Usage	Low (read from disk)	High (persistent server processes)	Medium (API calls)
Privacy	100% local	100% local	Requires external LLM APIs
Semantic Search	Local Ollama or keywords	❌	OpenAI/Anthropic embeddings
Call Graph	Bidirectional with alias resolution	LSP-based	❌

When to choose CICADA: You want local-first operation with rich git context (PR attribution, blame, function evolution tracking) and efficient token usage.

When to choose Serena: You need code editing capabilities through LSP and can accept higher resource usage.

When to choose Codicil: You have an Elixir project and prefer LLM-powered semantic summaries (Elixir-only).

---

Contributing

git clone https://github.com/wende/cicada.git
cd cicada
uv sync
pytest

Before submitting a PR:

• Run black cicada tests
• Ensure tests + coverage pass (pytest --cov=cicada --cov-report=term-missing)
• Update docs if behaviour changes

We welcome issues/PRs for:

• New language grammars
• Tool output improvements
• Better onboarding docs and tutorials

---

License

MIT – see LICENSE.

<div align="center">

Stop wasting context on blind searches. Give your AI CICADA.

Get Started · Report Issues

</div>