AI assistants and MCP
How AI assistants use done. through the MCP server, with sign-in instead of keys, and what they always ask first.
This page explains how an AI assistant uses done. through the MCP server, how it signs in, and the safeguards that always apply.
MCP is a standard way for an AI assistant to use tools. The done. MCP server gives an assistant a small set of tools built around jobs, not around endpoints. It is generated from the same description as the API, so a tool never does anything the API would not.
Preview. The MCP server opens with the API. Nothing here is live yet.
Add the server
Add done. to your assistant as a remote MCP server:
| Setting | Value |
|---|---|
| Address | https://api.done.com/mcp |
| Transport | Streamable HTTP |
| Sign-in | OAuth. The assistant opens a browser so the person can sign in. No key is pasted. |
Sign in, no keys
An assistant does not use a pasted key. The person signs in with done., approves access in a browser, and picks the workspaces the assistant may touch. They can approve read, spend or full. See Keys, workspaces and sign-in for the scopes and the sign-in addresses.
Sends made through sign-in are recorded as API sends. They are never recorded as the signed-in person's own reply, so sending hours and limits always apply.
The tools
The server offers about 12 tools. Each tool covers one job and uses the API operations behind it.
| Tool | The job |
|---|---|
what_needs_me | See what needs you today, and what is stuck or failed. |
find_people | Search LinkedIn for people (full plan). |
check_fit | Score one person or a list, preview a run, and follow its progress. |
research_company | Research a company and build a customer profile. |
scouts | Create, run and read scouts, and decide their first look and review. |
campaigns | Build, launch, pause and read campaigns (full plan). |
add_to_campaign | Add people to a campaign, or send a scout's people to one. |
inbox | Read conversations and messages, and send a reply (full plan). |
draft_reply | Write a reply for the person to approve (full plan). |
approvals | List drafts, and approve, reject, edit or rewrite them (full plan). |
results | Read results for a scoring run, a campaign or the whole workspace. |
goals | Propose a goal, then approve, start or stop it (full plan). |
Tools marked full plan are not part of the API-only plan. See the Introduction.
What an assistant asks first
Reading needs no question. Anything that spends, sends or changes asks the person first.
- Actions that spend credits, such as scoring a list or researching a company, ask first.
- Actions that send, or that can lead to a message reaching a real person, such as launching a campaign or approving a draft, ask first.
- Actions that change your setup, such as editing a campaign or deleting a list, ask first.
The permission the person approved still limits what is possible. A read sign-in cannot spend or send, whatever the assistant is asked.
The first look is always on
A scout made through the MCP server always has its first look on. The new scout finds and checks 20 people, then stops and waits for the person to decide what happens next: keep going, look again, or stop. No one is handed to a list or a campaign before the person has looked at the first results.
first_look is required when a scout is created, so callers of the plain API send true or false. The MCP server always sends true. With false, people_cap is required: the most new people one run may find.
What it never does
An assistant is held to the same rules as every other caller. See What done. never does through the API. It never sends to LinkedIn outside a campaign or a reply, never skips an account's pacing, limits or sending hours, never automates replies to people who raised their hand, and never shows how a score is worked out.
Using the API directly
An assistant can also call the API with a key. A person gives it a key made in Settings, API keys, and the assistant reads the API reference. The same description is available as plain text for models:
/llms.txtlists the guides, each with its Markdown copy./llms-full.txtholds every guide in one file, in navigation order./openapi.jsonis the full API description.- Every guide is also available as Markdown at
/llms.mdx/docs/<slug>/content.md, where<slug>is the guide's path after/docs/. For example,/llms.mdx/docs/quickstart/content.mdand/llms.mdx/docs/building-blocks/scoring/content.md. The Introduction is/llms.mdx/docs/content.md. The copy button at the top of each guide gives the same text.