Running LinkedIn outreach from an AI agent with the Slicegraph MCP

An outreach campaign is a handful of decisions: who, in what order, with which words, under which limits. Those are exactly the decisions an agent is good at drafting and bad at being trusted with. The Slicegraph MCP server is built around that split. An agent can do everything up to the point where a real account starts sending; a signed-in person does that part.

What MCP is, in one paragraph

The Model Context Protocol is the standard way to hand an AI model a set of tools. A client such as Claude connects to a server over HTTP, reads the list of tools with their descriptions and argument schemas, and calls them during a conversation. Slicegraph exposes one at /api/mcp, authenticated with a bearer key you create in the console.

The tools

Tool	What it does
list_senders	Connected LinkedIn senders with today's sampled limits, warmup and cooldown state
list_campaigns, get_campaign	Campaigns with counts; one campaign's settings, sources, workflow and readiness
create_campaign, update_campaign	Draft a campaign, rename it, pick a sender
set_campaign_cadence	Choose a safety preset or slow a campaign below the sender's own limits
apply_workflow_template, set_campaign_workflow	Start from a built-in template, or write the full step graph
search_filter_options	Resolve a typed value (a title, a location, an industry) to LinkedIn's option ids
start_audience_preview, get_audience_preview, import_audience_preview	Stage a search, read the 250-row preview with dedupe verdicts, import the rows you want
add_campaign_source, set_campaign_source_active, detach_campaign_source	Attach a pasted URL, a search or a signal feed; pause or remove it
pause_campaign	Stop a running campaign on the worker's next pass

Eighteen tools in total. The full reference with argument schemas is on the documentation page.

A session, end to end

A realistic conversation with an agent looks like this:

1. "List my senders." The agent sees one connected account, on Balanced, warmup complete.
2. "Draft a campaign for accountants in Antwerp who just registered on Peppol." The agent creates the campaign, resolves "Antwerp" and "accountant" through the filter typeahead, and stages a preview.
3. "How many?" The agent reads the preview: LinkedIn's own estimated total plus the 250-row sample with new, duplicate and recently-contacted counts.
4. "Use the signal-led template and import the new ones." The agent applies the template and imports.
5. "Check it is ready." The agent calls get_campaign and reads the readiness result: sender connected, source active, no workflow errors.

Then it stops, because the sixth step is not available to it.

What an agent cannot do

An MCP key can never activate a campaign. Activation is the moment a real LinkedIn account starts sending on someone's behalf, and that click stays with a signed-in person in the console. The same readiness check runs there: no connected sender, no active source, or a template with a leftover [placeholder] refuses to activate and says why. An agent can pause a campaign, which is the safe direction, and cannot resume one.

The other boundary is the sender itself. Every tool call runs under the sender's caps, warmup and cooldowns exactly as the console does; there is no privileged path. The sending limits post has the figures.

Where to start

Create an MCP key on the accounts page, add the server to your client, and ask it to list senders. The outreach page describes what the campaigns do once they run, pricing is €29 per sender per month, and if the audience should come from a company event rather than a search, the tender feed for Amsterdam is a good first source to point the agent at.