Devenv Cache

AI-Native Local Dependency and Documentation Context Cache via Model Context Protocol (MCP)

Built with help from Antigravity, Google DeepMind's agentic AI coding assistant.

DevEnv-Cache is a lightweight, zero-dependency developer tool designed to bridge the gap between AI coding assistants (like Cursor, Claude Code, Copilot, and Antigravity) and your project's local dependencies. It parses project lockfiles, builds a local SQLite full-text search database of package READMEs/APIs, and serves this context offline instantly via an MCP server.

---

🔴 The Problem

AI coding assistants are highly capable but suffer from critical friction points when working with external packages:

1. API Version Hallucinations: Assistants frequently hallucinate outdated or incorrect API endpoints because their training data is static (e.g., mixing Pydantic V1 and Pydantic V2 signatures).
2. High Latency & Slow Iteration: When assistants crawl documentation or perform web searches on the fly to find library details, it takes 1.5 to 5.0 seconds per query.
3. Token Pollution: Scraping full web pages injects thousands of noise tokens (headers, footers, auth links) into the prompt context, raising costs and risking context truncation.
4. Offline Blockers: If you are coding on a plane, in a secure sandbox, or have intermittent internet, assistants lose the ability to lookup documentation entirely.

---

🟢 The Solution

DevEnv-Cache indexes documentation locally, making it accessible to AI agents in under 10 milliseconds with zero internet roundtrips.

graph TD
    A[Lockfiles: poetry.lock, requirements.txt] --> B(devenv-cache sync)
    B --> C{Docs Fetcher}
    C -->|Markdown Parsing| D[(Local SQLite FTS5 DB)]
    E[AI Agent / IDE] -->|MCP Request| F[MCP Server]
    F -->|Instant Query| D

---

🛠️ How to Use It Effectively

1. Installation

Install the CLI tool inside your virtual environment (or globally):

pip install .
# Or using uv:
uv pip install -e .

2. Initialize the Cache

Navigate to your Python project's root folder (containing poetry.lock or requirements.txt) and run:

devenv-cache init

This creates a .devenv-cache/ configuration directory and sets up the local SQLite database.

3. Sync Dependencies

To parse your lockfile and cache the documentation from PyPI:

devenv-cache sync

• Best Practice: Run this command whenever you add, remove, or update dependencies in your project (e.g. after running poetry update or pip install).

4. Register the MCP Server

To allow your AI coding assistant to query the cache, register it in your assistant's configuration file (e.g., mcp_config.json globally at ~/.gemini/config/mcp_config.json or your IDE settings):

{
  "mcpServers": {
    "devenv-cache": {
      "command": "devenv-cache",
      "args": ["mcp"]
    }
  }
}

---

💡 Developer Best Practices: How to Code Effectively

When DevEnv-Cache is active, you don't need to change your coding habits or write complex prompts. The system works transparently:

• Write Prompts Naturally: Prompt your agent as you normally would: "Write a custom FastAPI middleware to track custom requests."
• Behind-the-scenes magic: The AI agent notices fastapi in the codebase or prompt, automatically invokes the devenv-cache MCP server tools (search_dependency_docs), and receives the exact API signatures matching the version installed on your machine.
• Keep it Offline: If you go offline, the agent will continue to write perfect imports and method signatures without throwing errors.

---

📊 Run Benchmarks & Tests

Verify Code Integrity

Run the comprehensive test suite to confirm all units (database, cacher, parser, mcp) are fully operational:

uv run pytest

Run Performance Comparison

To see the comparison report between coding with and without DevEnv-Cache:

uv run python test_bed/run_benchmark.py

Open docs/benchmark_report.md to inspect the latency differences (e.g. 0.0004s local SQLite search vs. 0.15s remote HTTP fetches) and token savings.

---

🤖 Built with Antigravity

This repository was designed, built, and tested in partnership with Antigravity, Google DeepMind's agentic AI coding assistant. The development process adhered strictly to Test-Driven Development (TDD) protocols, resulting in a clean, zero-dependency implementation of the MCP standard.