LinkedIn MCP tools

This page and the endpoint read from the same array in the codebase, so the reference cannot list a tool the server does not implement. If a capability is missing here, it is missing there — see what is not exposed for the things people most often assume are present. Setup is on the Claude page; the LinkedIn accounts these tools drive are covered under accounts for agents.

Two surfaces, and why the difference matters

Every tool falls on one side of a line, and it is the first thing to understand about this server because it governs cost, risk and rate limits alike.

LinkedIn tools

These are the reason to connect this server rather than read a dashboard. Each one runs as a named sender account: pass account to choose, or leave it out and the first connected account is used.

Workspace tools

Your own outreach data, and the accounts the LinkedIn tools run through.

How to call a tool

Every call is the same envelope; only params changes.

POST https://topclozer.com/api/mcp
Authorization: Bearer tcz_live_YOUR_KEY
Content-Type: application/json

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": { "name": "<tool>", "arguments": { } }
}

Results come back as MCP content blocks — one text block whose body is the tool's JSON, pretty-printed. Parse the string; there is no structured-content field.

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "content": [ { "type": "text", "text": "<the tool's JSON, pretty-printed>" } ]
  }
}

Errors

Failures split by who can do something about them. A problem the calling agent can act on — no connected account, an unresolvable profile URL, a sender that is paused — comes back as a successful response carrying isError: true and a sentence explaining what to do. That is what lets an assistant read the message and try something else instead of reporting a transport failure to the user.

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "isError": true,
    "content": [ { "type": "text", "text": "This workspace has no connected LinkedIn account…" } ]
  }
}

Genuine protocol faults stay JSON-RPC errors:

Note the asymmetry: only auth and the rate limit carry a non-200 status. Check the body, not just the status code.

Rate limits and pacing

30 LinkedIn tool calls per 5 minutes, counted per workspace. Shared across every LinkedIn tool, and per workspace rather than per key — a second key does not buy a second allowance. A single search_linkedin_people call returns up to 50 profiles, so the budget goes further than it looks.

It is there for your benefit, not ours. An agent retrying in a loop is precisely the request pattern that gets a LinkedIn account restricted, and a 429 you can retry past is a far better outcome than an account you cannot. Workspace tools are unlimited.

The other limits live inside the product and govern sending, which no tool here can trigger: a per-sender weekly invite cap, a warm-up ramp that starts a new account low and steps it up, a working-hours window evaluated in the workspace timezone, and a randomised gap between two actions on the same sender. They apply identically whether a person clicks send or an agent triggers the workflow.

What is not exposed, and why

• Sending anything — no invite, message, reply, follow or like tool. Outbound actions are released by the scheduler under the workspace review mode, which by default queues every invite and message for a person to approve. This is the boundary the whole product is arranged around, not an unfinished feature.
• Approving a draft — list_tasks shows the queue so an agent can summarise it. Clearing it is a person's job.
• Writing leads or campaigns — an agent can search LinkedIn and read the result, but importing those people into a campaign happens in the app. pause_campaign is the only write, and deliberately one-way.
• Company and job search — people search is implemented and proven; company and job entities are not, and we would rather have four LinkedIn tools that work than seven where three are guesses.
• Account management — connecting an account, changing proxies, editing sending limits. Not implemented, and not planned as an agent surface.
• MCP resources and prompts — the server advertises the tools capability only.
• Paging and since-timestamps — list_leads and get_inbox return the full set. Filter client-side.

Changelog