What is an MCP harness?

An MCP harness is the set of tools, stored state and guardrails that an AI agent works through to do a job that outlasts a single conversation. MCP (Model Context Protocol) is how the agent connects to it; the harness is what sits on the other side.

The agent brings the reasoning: reading, deciding, writing. The harness brings what reasoning alone cannot: memory between sessions, a clock that keeps running when the agent is not, and limits on the actions that cannot be undone.

Why "harness"

Agent builders already use the word for the scaffolding around a model: the loop, the tools and the context it runs inside. The plain meaning fits too. A harness is what lets a strong animal pull a load in a useful direction without hurting anyone.

What a harness contains

• A system of record. State the agent reads and writes: in outbound, accounts, contacts, research and history.
• Tools shaped around the job, not raw APIs. A few task-sized tools with compact results, so the agent’s context goes to the work rather than to tool definitions.
• Time. Things that happen while the agent is not running: scheduled steps, waits, and wake-ups when an event arrives.
• Guardrails on irreversible actions. Daily limits, fixed recipients, one reply per message, and checks before anything is sent.
• An audit trail. Evidence and history, so a person can see why the agent decided what it did.

Connector, harness or AI SDR?

• A connector exposes one app to the agent: send an email, read a CRM record. It holds no state of its own and knows nothing about the job.
• A harness is shaped around a job and holds the state between steps. It usually works alongside connectors.
• An AI SDR platform is a finished product with its own agent inside. You hand the job over instead of running it with your own agent.

Why outbound needs one

Outbound has every property that breaks a bare agent:

• It runs for weeks, far longer than any session.
• It is timed: cadences, waits and the recipient’s working hours.
• It is irreversible: a sent email cannot be unsent, and a message to the wrong person on LinkedIn cannot be taken back.
• It is event-driven: a reply should change what happens next, within minutes.
• It is high-volume: hundreds of accounts do not fit in a context window as raw data.

Harp

Harp is the MCP harness for pipeline generation. Your agent researches, writes and decides through Harp’s MCP tools; Harp holds the records, runs sequences across email and LinkedIn, sends from your own mailbox and wakes your agent when someone replies. It is set up with Claude today. Because the agent is yours, your outbound improves as soon as the models do. Start with How to run outbound from Claude.