cachly-mcp

Purpose

cachly provides persistent memory for programming AI assistants, solving the problem that when the same issue arises repeatedly, the AI cannot remember past experience. It connects editors to AI assistants via the MCP protocol, stores fixes, technical decisions, and failed lessons, and automatically recalls them when needed.

Main Capabilities

• Lesson storage: Each fix, decision, or dead end is stored as a lesson, containing topic, what worked, what did not work, severity, file paths, commands, and author.
• Session briefing: At the start of a session, the AI receives a briefing containing the last session, unresolved failures, and recent lessons.
• Smart recall: smart_recall combines BM25 keyword search, semantic-based search, and a causal graph, supporting 11 languages.
• Causal tracing: causal_trace starts from error symptoms and follows the causal knowledge graph to find root causes, intermediate causes, and effective fixes.
• Failure prediction: brain_predict lists failure patterns from past incidents before deployment.
• Team sharing: Teams can share a Brain, and lessons carry the name of the learner.
• CI integration: Learn from commits and red-green transitions via Git hooks and CI templates.
• 124 MCP tools: 27 are listed by default (26 everyday tools + the dispatcher cachly_tool), and the remaining 98 are invoked through the dispatcher.

How to Use

• Start without an account: With no key, a test Brain is automatically created on the EU server at the start of a session, providing a 14-day Dev tier experience.
• Automatic configuration: Run npx @cachly-dev/mcp-server@latest autopilot to detect and configure common editors, add Git hooks to learn from commits, and log in via a browser device code.
• Manual addition: You can manually configure the MCP server; with no key, the first tool call creates a test Brain.
• VS Code extension: Install the cachly Brain extension, and a Brain is created on first launch.
• JetBrains plugin: Install from the JetBrains Marketplace, supporting IntelliJ IDEA, PyCharm, GoLand, WebStorm, and Rider.
• Demo: Run npx @cachly-dev/mcp-server@latest demo to read Git history in the project folder and print fixes, patterns, and decisions; data never leaves the local machine.
• Data export:npx @cachly-dev/mcp-server@latest export exports all lessons as lessons.md and lessons.jsonl.
• Environment variables: CACHLY_ALLE_WERKZEUGE=1 lists all 124 tools; CACHLY_PROFILE=lean lists 8 plus cachly_tool; CACHLY_PROFILE=recall lists 3 read-only tools.

Use Cases

• Individual developers: Avoid solving the same problem repeatedly and build a personal knowledge base.
• Team collaboration: Share team experience so new members can reuse existing knowledge.
• CI/CD pipelines: Learn from every CI run; red-to-green indicates a verified fix, and green-to-red indicates a known cause.
• Self-hosting: Can point to a private backend and supports bring-your-own embedding computation.

Limitations

• By default, embedding computation takes place on cachly's EU servers; text is sent to a third party only when the user specifies a third-party provider.
• API traffic passes through Cloudflare (a US company).
• Code snippets are stored only when the user actively calls index_project, and only on the user's own EU instance; by default, up to 1200 characters of summary are stored per indexed file.
• Other tools store lesson text, commit messages, session summaries, and key-value context, but do not store source code.
• The free tier has a single-Brain limit.
• Current pricing and limits are available at cachly.dev/pricing.