Connect Signl as an MCP server and Claude can run the whole engagement workflow in conversation: add a source to watch, read the people it found and why they fit, rewrite an opener, and push the right ones into your sequencer. The agent calls named tools; it never drives your LinkedIn account.
Disclosure: Signl is ours. The tool names below are the ones documented in our MCP tool reference and checked against apps/api/src/mcp/server.ts on 24 October 2026. If you want the older company and buying-signal side of the server instead, that is connect Claude to B2B data.
Key facts
- Claude calls tools on a remote server. It does not browse, scrape or log into LinkedIn as you.
- Tool calls are free. Signl is priced by sources watched, so an agent reading the inbox repeatedly costs nothing.
- Anything outbound or irreversible is confirmed with you before it happens.
- The agent can do what the app does, including setting up the ICP and the autopilot rules.
- Sending stays in your own outreach tool. Signl prepares, qualifies and routes.
Why an agent is a good fit for this particular job
Engagement prospecting is a lot of small judgements made repeatedly: is this person a buyer or an agency, does this comment carry a real problem, is this opener's quote actually in the comment, should this one go to a campaign or a human. Those judgements are the work, and they are exactly what a conversation is good at and a dashboard is bad at.
The part that makes it practical is that the expensive operations are already done server side before the agent sees anything. By the time Claude calls get_inbox, the people in it have been collected, deduplicated, filtered for noise, judged against your ICP, enriched where they passed, and given a drafted opener. The agent is reasoning over finished work, not orchestrating six API calls and hoping.
Connecting
Two minutes, and it is the same connection for every workflow below.
- In Claude, open Settings, then Connectors, then Add connector.
- Paste the Signl MCP server URL shown under Settings, then Agent, in the Signl app.
- Approve access. You sign in to Signl and confirm; no secret is copied anywhere.
An API key beginning sk_signl_ is the alternative for clients without an OAuth flow, including the CLI. Keys are shown once and stored as a SHA-256 hash, so a lost key is replaced rather than recovered. The full walkthrough, including what to do when the client shows no tools, is in connect your agent.
Then confirm it worked by asking the obvious question:
What Signl tools do you have access to?The tools that matter, by job
You never call these yourself. Knowing the names is still the fastest way to phrase a request the agent can act on.
| Job | Tools | What you would say |
|---|---|---|
| Set up who you sell to | get_icp, analyse_website, update_icp | "Read our website and draft an ICP, then show me before saving" |
| Add and manage sources | watch_posts, suggest_sources, list_watches, update_source, stop_watching | "Watch our two closest competitors' founders, comments only" |
| Work the inbox | get_inbox, get_person | "Who engaged since yesterday that fits, and why" |
| Fix the first message | edit_opener | "That opener names the post. Rewrite it answering the reconciliation line" |
| Triage | dismiss_people, restore_people, mark_contacted | "Dismiss the job seekers and tell me the reason you used" |
| See the account view | list_companies, get_company | "Any company with three or more people engaged this month" |
| Route onward | list_destinations, connect_destination, list_campaigns, send_people | "Send the four best to my Instantly campaign" |
| Automate it | get_autopilot, set_autopilot, delete_autopilot | "Rule: our own posts only, confidence 70 and up, max 10 a day" |
| Account and outcomes | get_account, get_pipeline_report, get_upgrade_link | "Which sources produced replies last month" |
| Privacy | record_objection | "This person asked not to be contacted. Suppress them" |
import_from_trigify also exists, which moves watched profiles and searches in from an export or a pasted list.
A worked morning
Four turns, in the order they actually happen. The agent's answers are paraphrased, not transcripts.
Turn 1, the ICP, once.
Read heysignl.com and draft an ICP for us. Show me the fields before you save anything.The agent calls analyse_website, proposes what you sell, the titles, seniorities, industries, countries and headcount bounds, and waits. Correct it in conversation rather than in a form, which is the part that is genuinely faster than the app. Ten minutes here decides the quality of everything after it, as define your ICP argues at length.
Turn 2, sources.
Watch these three: our founder's profile, Acme's founder's profile as a competitor,
and a topic search for "month end close" and "reconciliation". Comments only
except on ours.One watch_posts call per source. The relationship matters and is not editable afterwards, because it governs what a drafted opener may say: on a competitor source, Signl rejects any opener naming the post's author or referring to their comment. A topic search counts as 2 sources against the plan's 25.
Then, a fortnight later, the question worth asking:
Which authors do our qualified people keep engaging with that we are not watching?That is suggest_sources, and it answers from your own collected data rather than a guess. The rules are in sources Signl finds for you.
Turn 3, the inbox.
Who engaged since I last asked, only the ones who fit, and show the comment for each.get_inbox with new_only. You get people ranked by priority, each with the comment, the source, likely intent, a reason in words and a drafted opener. Then the follow-ups a conversation makes cheap:
Of those, which wrote about a specific number or deadline?
Show me the two you are least sure about and why.The second question is the one to build a habit around. Reading the uncertain cases is how you find out your ICP is wrong, and dismiss_people takes a reason that feeds back into ranking.
Turn 4, routing.
Send the three strongest to my Instantly campaign, and draft a public comment
reply for each instead of a message where the comment deserves one.list_campaigns then send_people, and the agent confirms before anything leaves. Nine destinations exist: Instantly, Smartlead, lemlist, HeyReach, Expandi, HubSpot, Clay, a generic webhook and Slack alerts. The field mapping each one receives is in sending engagement leads to HubSpot and Clay.
Autopilot, carefully
The agent can write routing rules, which is the point at which you should slow down.
A rule says: people at or above this confidence, from these source relationships or source kinds, go to this campaign, at most this many a day. An account can hold 10 rules and sends at most 500 people a day across all of them. Rules run after each collection and never send the same person to the same campaign twice.
Our advice, as the people who built it: run the first month with no rules. Read twenty drafted openers yourself. Then add one rule, for your own posts only, at a high bar. Competitor sources last, if at all, because that is the source where a bad message costs the most.
What the agent will not do
- Send a LinkedIn message. Native sending from your own account is listed as coming soon on our pricing page. Today, LinkedIn delivery happens in your own HeyReach or Expandi account.
- Create or revoke API keys, connect Slack, or disconnect a tool. Those stay in Settings, on screen, because each one grants or removes access to your account.
- Act irreversibly without asking. Sending people, removing a source, deleting a rule and recording an objection are all confirmed first.
- Reveal a credential. Keys and webhook URLs you paste are stored server side, never repeated back, and never kept in Signl's own call log.
Limits
- Client support varies. The server speaks MCP over HTTP with OAuth or a bearer key, and connector behaviour differs between Claude, ChatGPT, Cursor and the rest; our read of the landscape is MCP clients compared.
- Large tool results get truncated by the client, not the server. This is a real failure mode in agent workflows and it is quiet when it happens; we wrote up the mechanics in why your prospecting agent silently drops results.
- The agent is only as good as the ICP. An agent with a vague ICP produces confident, wrong qualification faster than a human would.
- Only public engagement is collected. No follower lists, no connections, no post viewers. Automated collection of public engagement is also contractually contested: LinkedIn's User Agreement section 8.2 prohibits scraping and bot access (read 24 October 2026), and the separate data protection question is in GDPR and B2B prospecting. This is not legal advice.
Next: LinkedIn social listening for the method the tools implement, or MCP for sales if you are still deciding whether an agent belongs in your stack at all.
