Observer v0.1.0 - transparent MCP proxy for agent observability

[tamish560]

We analyzed 1665 discussions across 8 major AI agent communities (MCP, Codex, Cline, Goose, Anthropic SDK, OpenAI Python, shadcn, agent-orchestrator) and found the #1 recurring problem is agent observability. Nobody can see what their agent is actually doing.

So we built Observer.

What it does:

Observer is a transparent proxy that sits between your MCP client (Claude Desktop, Cline, Goose, Codex) and the actual MCP server. It:

• Logs every tool call to SQLite (input, output, duration, error status, token estimate)
• Exposes 4 trace tools to your agent: trace.history, trace.stats, trace.search, trace.replay
• Optionally filters tools to reduce token overhead (the implement tool filtering to reduce token overhead #2 problem found in our research)

Quick start:

OBSERVER_TARGET="npx -y @modelcontextprotocol/server-filesystem /tmp" observer

That's it. Your MCP client now talks to Observer, which proxies to the real server and logs everything.

What's next:

• SSE transport support (add support for SSE transport #3)
• OpenTelemetry export
• Tool recommendation engine (semantic filtering based on context)
• npm wrapper for JS/TS ecosystem

We have 2 good first issues open if you want to contribute:

• add test coverage for store package #1 Add test coverage for store package
• implement tool filtering to reduce token overhead #2 Implement tool filtering to reduce token overhead

The codebase is small Go + SQLite. No frameworks, no external dependencies beyond go-sqlite3. Read CONTRIBUTING.md for the AI agent contribution guide if you use AI coding tools.

Feedback welcome. What trace tools would be most useful to you?