Operations agents for LinkedIn

How many LinkedIn tool calls this workspace has made this month, broken down by tool, plus the rate limit in force. The call an agent makes to pace itself instead of discovering the limit by hitting it.

Runs on get_usage · Read only

Reads sender status, warm-up stage, daily invite and message counters and the health score, and turns them into a decision with a reason: keep going, slow down, or stop. Written down because the failure it prevents — a restricted LinkedIn account — is not reversible by changing a setting afterwards, and because the signals that precede it are visible in what these tools already return.

Performed by a skill — instructions your AI client follows. · Read only

Turns a piece of work into an arithmetic question against three separate budgets — thirty LinkedIn calls per five minutes per workspace, the monthly allowance the usage report names, and each sender's daily caps — and says what fits, what has to be paced, and what does not fit at all. The point is to discover a limit on paper rather than by hitting it mid-run.

Performed by a skill — instructions your AI client follows. · Read only

Every campaign in the workspace with its live funnel counters — invited, connected, replied — and the two rates already rounded to whole percentages.

Runs on list_campaigns · Read only

The AI-written replies waiting for a person to press send. An agent can read and summarise this queue; it cannot approve anything in it, and that is deliberate — sending is the one irreversible action in the product.

Runs on list_tasks · Read only

The workspace balance and the entries behind it, with whether the balance can refuse anything at all reported alongside the number — because enforcement is off by default and a meter read as a spending limit is worse than no meter.

Runs on check_credits · Read only

One thread per lead across every sender account, with full message bodies and — once a reply has been classified — its intent. There is no since or limit argument; the whole inbox comes back.

Runs on get_inbox · Read only

Senders, campaigns, leads, invited, connected and replied, counted across the whole workspace. The cheapest way to answer "how is outreach going" without walking every campaign.

Runs on get_stats · Read only

Installs, purchases, the hosted seat, credits and the execution ledger over a window: what ran, what succeeded, what failed and the class of failure. The same rows the Executions page shows a person.

Runs on get_platform_usage · Read only

What has been installed, where, at which version, and what was paid for it — including removed installs when asked for the full history. The library, read back as rows rather than as a page.

Runs on list_installations · Read only

The accounts this workspace can act as, each with status, warm-up stage, how much of today's invite and message budget is spent, and a health score. Reads our own database, so it costs nothing against the LinkedIn rate limit and is the cheapest first call an agent can make.

Runs on list_sender_accounts · Read only

The only tool that writes anything, and it writes to our database rather than to LinkedIn: it stops outbound. There is no resume tool, because restarting a campaign is a deliberate human action in the app.

Runs on pause_campaign · Read only

A fixed pass over the things that go wrong quietly: sender health and checkpoints, how much of each daily cap is spent, the monthly LinkedIn allowance against the calendar, the approval queue's age, the execution ledger's failures by class, and the credit balance. It ends in a written decision per account and, where the answer is stop, the one tool call that stops it.

Performed by a workflow — a sequence of steps, not a single call. · Human approves