The LinkedIn API for AI agents — with limits built in.
One MCP integration for inbox, messages, connections and feed. Linbox enforces safe rate limits so your agent can't get the account banned.
works with Claude · Cursor · OpenAI Agents SDK · any MCP client
{
"status": "queued",
"task_id": "tsk_1a2b3c",
"scheduled_for": "2026-07-06T14:20:00Z",
"data_source": "cache",
"as_of": "2026-07-06T13:58:11Z"
}→ no error, no ban. The agent keeps working and polls linbox_get_task later.
Agents + LinkedIn = ban risk.
Linbox is the safe layer between your agent and LinkedIn — it does the automation and keeps the account alive.
- 01
LinkedIn has no official API for messaging or connections. Agents scrape.
- 02
Raw automation ignores rate limits → LinkedIn flags it → the account gets restricted or banned.
- 03
Existing tools automate actions but don't manage limits. That's the gap.
Connect once. Let the agent work.
The Chrome extension captures the session. Linbox runs a server-side browser with a per-account proxy.
Plug the endpoint into your agent with a Bearer token. Works with any MCP client.
Read the inbox, reply, send and accept connections, tag people — all through typed tools.
Over the limit? The action becomes a task. The agent returns later with a task_id.
One integration, the whole LinkedIn surface.
List threads, read full history, reply. Incremental sync keeps everything current.
Send, accept and withdraw invites. Dedup plus friend detection built in.
Search the local DB, tag and filter people and threads.
Parse public profiles — headline, company, connections, degree.
Organize relationships with tags the agent can read and write.
Read and act on the feed. Shipping soon.
Built-in rate limits.
Every action type has its own budget — per minute, hour, day, week and month — plus a minimum interval and the account's work hours and timezone. Ship a conservative or aggressive preset, or bring your own config.
Configurable per account. Never exceeded.
| Action | /hour | /day | /week |
|---|---|---|---|
| Send message | 5 | 20 | 80 |
| Connection request | 10 | 40 | 200 |
| Read inbox | 4 | 12 | — |
numbers = linkedin_high preset — illustrative
Over the limit? The agent gets a task, not an error.
When an action can't run now, Linbox queues it and returns a task_id plus scheduled_for. The agent keeps working and polls linbox_get_task later. Batches get a batch_id. Nothing is lost, nothing blocks.
The request fails, the run dies, and repeated retries push the account toward a ban.
The agent gets a task_id, moves on, and picks up the result later. No dead ends.
Designed for agents, not dashboards.
MCP-native
Standard tools with typed contracts and JSON or markdown responses. Drop them into any MCP client — no SDK lock-in.
Always a usable answer
Reads fall back to cached DB data with data_source and as_of, so the agent knows how fresh it is and why live didn't run.
What people build on Linbox.
Qualify, connect and follow up within safe limits — automatically.
Triage threads and draft replies with full conversation context.
Pull relationship context — messages, connections — into your own app.
Add LinkedIn as a tool to an existing agent in minutes.
Works with your stack.
Authenticate with a Bearer token, or pass ?token=for clients that can't send custom headers.
Give your agent LinkedIn — safely.
Connect an account, grab a token, start calling tools.