agentwatch MCP server

Expose your local agent history to any MCP-compatible client (Claude Code, Cursor, custom LangChain/CrewAI agents, …) so running agents can inspect what they — or other agents — did before. All local. No network. No telemetry.

Contract

GOAL: Serve this machine's agent history via MCP so any client can query it without cloud or custom wiring. USER_VALUE: A running agent can ask "what did I do last session?" and hand context to the next agent without manual copy-paste. COUNTERFACTUAL: Agents repeatedly re-discover what a prior run already learned; no cross-agent handoff across the multi-agent stack.

Quick start

# Option A — if agentwatch is on your PATH (npm i -g)
claude mcp add agentwatch -- agentwatch mcp

# Option B — npx on demand
claude mcp add agentwatch -- npx -y @misha_misha/agentwatch mcp

Restart Claude Code, then run /mcp inside the TUI — agentwatch should appear with 5 tools listed.

Manual install (edit config by hand)

Claude Code

Edit ~/.claude.json (or ~/.claude/config.json depending on your version):

{
  "mcpServers": {
    "agentwatch": {
      "command": "npx",
      "args": ["-y", "@misha_misha/agentwatch", "mcp"]
    }
  }
}

Cursor

Edit ~/.cursor/mcp.json (or use Cursor's "Add MCP Server" UI):

{
  "mcpServers": {
    "agentwatch": {
      "command": "npx",
      "args": ["-y", "@misha_misha/agentwatch", "mcp"]
    }
  }
}

Generic MCP client (stdio)

Any client that speaks MCP over stdio. Example:

{
  "command": "npx",
  "args": ["-y", "@misha_misha/agentwatch", "mcp"]
}

Tools exposed

Tool	Args	Returns
list_recent_sessions	limit?: 1-100	[{agent, sessionId, project, lastActivity, sizeBytes}] newest first
get_session_events	sessionId: string, maxBytes?: 1024-10_000_000	Raw JSONL lines for that session (tail-capped)
search_sessions	query: string, limit?: 1-50	[{session, agent, line}] substring hits
get_tool_usage_stats	sessionId?: string, limit?: 1-500	{scannedSessions, turns, tools: [{tool, count, totalDurationMs, errorCount}]}
get_session_cost	sessionId: string	{totalCostUsd, turns, tokens:{input,cacheRead,cacheCreate,output}, byModel}

Example agent prompts

After wiring it up, try asking your agent:

• "Use agentwatch to list my five most recent sessions and summarize what I was working on."
• "Use agentwatch get_tool_usage_stats to tell me which tools Claude has been failing on most often this week."
• "How much have I spent on session <id> today? Use agentwatch."

Notes

• All local. The server reads ~/.claude/projects/**/*.jsonl and ~/.codex/sessions/**/*.jsonl directly; no data leaves your machine.
• Fresh every request. Unlike the TUI, the MCP server does not maintain in-memory state — each call re-reads from disk. That means edits to sessions are reflected immediately, but heavy bulk queries (get_tool_usage_stats without sessionId) scan up to limit session files.
• No write tools. The server is read-only by design — an agent can look but can't modify session history.