An MCP server built for agents

Most MCP servers wrap an API: one tool per endpoint, verbose results, and a tool list that eats the agent’s context before it does any work. Harp’s server is shaped around the outbound job instead.

A small daily tool list

• briefing starts the session: what is due, what is held, what research is stale.
• find_accounts, account_brief and research_gaps find and inspect accounts.
• save_answers, upsert_accounts and upsert_contacts write research and records in bulk.
• playbooks and enroll start campaigns and enroll contacts.
• copy_review_queue and save_copy_reviews review drafts before they send.
• work_queue and step_done handle manual steps.
• pending_replies, reply_email, book_meeting and resolve_reply handle replies.

Folders for everything else

Less frequent tools live in folders, reached through three folder tools: browse_tools, describe_tool and run_tool. They cost nothing until your agent needs them.

Compact results

Results come back as compact JSON, with empty values dropped and dates shortened, so your agent can work through a full account list without filling its context window.

Where it runs

Harp’s MCP server runs locally next to Claude Code, Claude Desktop or Cowork, and talks to Harp over HTTPS with your API key. The Harp team installs it with you during setup.