@pinagent/mcp

stdio MCP server that exposes Pinagent feedback to a coding agent.

Install

npx @pinagent/mcp
# or with an explicit root
PINAGENT_PROJECT_ROOT=/path/to/repo npx @pinagent/mcp

The server walks up from cwd looking for .pinagent/ (then package.json) if PINAGENT_PROJECT_ROOT is unset, and logs the resolved root on startup.

Claude Code config

{
  "mcpServers": {
    "pinagent": {
      "command": "npx",
      "args": ["@pinagent/mcp"],
      "env": { "PINAGENT_PROJECT_ROOT": "${workspaceFolder}" }
    }
  }
}

Tools

| Tool | Purpose | | ----------------------------- | ---------------------------------------------------------------------- | | list_pending_feedback | List pending items, optionally filtered by since / file. | | get_feedback | Full record incl. screenshot as an image content block. | | resolve_feedback | Set status to fixed / wontfix / deferred (or back to pending). | | get_source_context | Return a numbered window of source lines around a file:line. | | get_conversation_transcript | Fetch the full persisted agent transcript for one conversation. |

get_conversation_transcript

Inputs: { id: string, format?: 'text' | 'json' }. Defaults to text — the same plain rendering pinagent transcript produces, so a spawned agent reading a prior run sees identical output regardless of which surface they came through. json returns the raw AgentEvent[] stringified for downstream parsing.

The transcript includes init / result events (the agent wants them for full context) and excludes the internal __finished bus sentinel. Reads directly from the .pinagent/db.sqlite messages table — no HTTP round-trip; same data the dock's transcript prefetch sees. Returns [] for invalid or unknown ids; 404-style "not found" errors only fire when the conversation row itself is missing.

Use it when an agent needs memory of its own prior runs (or another agent's) — for example, to reason about a follow-up turn after the user re-opens a landed conversation.

Channel mode (push events into a running Claude Code session)

The server also declares the claude/channel capability (research preview, requires Claude Code v2.1.80 or later). When Claude Code is launched with the channel flag, the server polls the local SQLite store and pushes a notifications/claude/channel event for each new comment so the agent reacts in your existing session instead of being polled.

Launch the session with both the project-scoped MCP config and the channel development flag:

cd your-project-with-pinagent
claude --dangerously-load-development-channels server:pinagent

The server:pinagent argument matches the key in your .mcp.json. Events arrive in Claude's context as:

<channel source="pinagent" id="..." file="src/Foo.tsx" line="42" col="7" ...>
  the developer's comment
</channel>

If you don't pass the flag, channel notifications are silently dropped and the pull-mode tools above continue to work normally.

Only comments left after the session starts are pushed. The watcher seeds itself with the IDs already in the store at boot and pushes events for items that arrive afterwards, so a backlog of comments queued before you launched the session won't trigger channel events. Reach those with the pull-mode tools (list_pending_feedback / get_feedback).