MCP server

Every coding agent starts each session with amnesia. It can’t see what you tried in a different tool yesterday, which approach you already rejected, or what you flagged as important.

Pounce can. It already reads sessions from Claude Code, Codex, opencode and Cursor on your machine, so it’s the one thing that can answer “what did Codex try on this bug yesterday?” — and pounce mcp hands that to any agent.

claude mcp add pounce -- npx use-pounce mcp

That’s it. The next time Claude Code needs context it can search everything you’ve done, in every agent, on this machine.

Setup

The MCP server talks to a running Bridge, so start one first:

npx use-pounce

Then register the server with your client. Installed globally (npm i -g use-pounce), use pounce in place of npx use-pounce everywhere below.

Claude Code

claude mcp add pounce -- npx use-pounce mcp

Or by hand — .mcp.json in your project root (shared with the repo), or ~/.claude.json for every project:

{
  "mcpServers": {
    "pounce": {
      "command": "npx",
      "args": ["use-pounce", "mcp"]
    }
  }
}

Claude Desktop

Settings → Developer → Edit Config, or edit directly:

• macOS — ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows — %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "pounce": {
      "command": "npx",
      "args": ["use-pounce", "mcp"]
    }
  }
}

Restart Claude Desktop completely afterwards — it only reads the config on launch.

Cursor

~/.cursor/mcp.json for every project, or .cursor/mcp.json for one:

{
  "mcpServers": {
    "pounce": {
      "command": "npx",
      "args": ["use-pounce", "mcp"]
    }
  }
}

VS Code (GitHub Copilot)

.vscode/mcp.json. Note the different shape — VS Code uses servers, not mcpServers, and wants an explicit type:

{
  "servers": {
    "pounce": {
      "type": "stdio",
      "command": "npx",
      "args": ["use-pounce", "mcp"]
    }
  }
}

Anything else

Most clients take the mcpServers shape above. If yours wants the command as one string, it’s npx use-pounce mcp.

Non-default port

If your Bridge isn’t on 8099, pass the port through:

{
  "mcpServers": {
    "pounce": {
      "command": "npx",
      "args": ["use-pounce", "mcp", "--port", "9000"]
    }
  }
}

The server finds the Bridge’s token itself over loopback, so there is nothing else to configure and no secret to paste into a config file.

Tools

Tool	What it answers
search_history	“Has anyone tried this before?” — full-text search across every agent
list_threads	“What sessions exist?” — newest first, filterable by agent or project
get_thread	“What actually happened in that session?” — the full message history
list_markers	“What did I flag as important?” — your own jump-to points in a thread
recent_activity	“What’s been worked on lately?” — activity across machines and repos

list_markers is the underrated one: markers are your judgement about what mattered in a long session, so they’re usually the fastest way for an agent to get oriented without reading thousands of messages.

Read-only, on purpose

The Bridge also exposes endpoints that run shell commands, start agent turns and make git commits. None of those are exposed over MCP.

Handing an arbitrary agent shell execution and commit rights through a tool call is a foot-gun, and it makes the security story impossible to explain in one sentence. Read-only keeps it simple: an MCP client connected to Pounce can read your history and nothing else.

Privacy

Everything stays on your machine. The MCP server runs locally over stdio and talks to the Bridge over loopback — no Pounce server sees your history, and nothing is uploaded. It’s the same data the Pounce app reads, exposed to a program already running on your computer.

Troubleshooting

no bridge on port 8099 — the Bridge isn’t running. Start it with npx use-pounce, then reconnect the MCP client.

Tools return nothing — check the Bridge sees your sessions first with pounce status, and that the agent whose history you expect has actually run on this machine. Pounce reads each agent’s own session files; it can’t see history that isn’t there.

Search returns no results — search is powered by ctx, installed to ~/.pounce/bin on first use. If it never installed, search_history reports that rather than returning empty.