What can an AI agent actually do on LinkedIn?
Search your own seat, import your own CSVs, research a person into a persistent column, run approval-gated sequences, send invitations and messages, comment and react on other people's posts, work the inbox, and report on all of it. It will not publish your posts and it does not sync anywhere.
Search and enrichment
Two sources of people, and no third. Search runs through your own LinkedIn seat, so the results reflect your entitlement and your real degrees of connection. Import takes any CSV you already own. External data providers were removed in September 2026 and are not coming back, so there is no bundled contact database anywhere in this.
Enrichment here is not a provider lookup. The agent reads a profile and writes what it found into a column on the lead group, where it persists, filters and sorts. The difference matters more than it sounds: a chat message about a prospect is gone next week, a column is still there when you build the next list.
The practical consequence is that list quality is a job you and the agent share rather than something you purchase. You choose the filter or the export. The agent reads each person and records what it found in a field you can sort on later. That is slower than buying a list, and it is why a list built this way survives the only question that matters when an account is scrutinised: why this specific person.
Sequences
Multi-step sequences with the send decision held back at every stage. Creating a sequence sends nothing. Enrolling a lead group queues drafts. Nothing leaves until a human approves it, and a reply stops the remaining follow-ups for that person immediately.
That last behaviour is the least glamorous thing in the product and it prevents the worst experience in outbound: the automated follow-up that lands after someone has already answered you.
Worth being precise about what a sequence is not. It is not a scheduler that runs unattended once configured. It is a queue of drafts with an enforced shape: the steps, the timing between them, and the stop condition. The send itself stays a decision a person makes, so a sequence across forty people is forty small decisions rather than one large one. That is the trade, and it is deliberate.
LinkedIn actions, including the one it will not do
Invitations, direct messages, InMail where your seat carries credits, profile views, comments on other people's posts, and reactions to recent posts. Sales Navigator Core carries 50 InMail credits a month if you hold a seat, and it does not raise the invitation limit, so InMail is a different lane rather than extra capacity.
It will not publish your own posts. That is not a hint at a roadmap, it is simply not in this build. The content capability is commenting and reacting on other people's posts, and nothing beyond that.
Every send passes the approval gate, and the caps are enforced server-side rather than typed in. A new seat runs at 20 invitations a day and 80 a week, multiplied by 0.5x in week zero, 0.75x in week one and 1.0x from week two. For context, LinkedIn itself publishes no invitation limit at all, daily or weekly. Our numbers are a policy we chose, not a ceiling we read off a page.
Inbox, learning and connect
Inbox: list conversations, read a thread, reply, tag, and stop a sequence to send something by hand when a thread deserves it. The point is that a good reply does not have to leave the place where the agent already has the history. Tagging matters more than it sounds, because it is how a qualitative read of a conversation becomes something you can filter a future list on.
Learning: knowledge you write once and the agent applies later, so your positioning, your disqualifiers and your tone are not retyped into every new chat.
Connect: link an account, check whether the link is alive, reconnect when LinkedIn invalidates a session, remove an account. Unglamorous, and everything else depends on it.
Observability
The agent activity log, outreach volume measured against the caps, per-account status, interaction history for a person or a company, and billing state. Anything the agent did, you can read back afterwards.
This category exists because an assistant acting on your behalf without a ledger is not something a careful person should run. The log is the difference between delegating and hoping.
It is also the only honest way to answer the question people ask after a week, which is what actually happened. Outreach volume measured against the caps tells you whether the ramp or your list is the current constraint. Interaction history for a company tells you whether a colleague already spoke to someone there. Per-account status tells you the link is alive before a sequence finds out that it is not.
What it cannot do
No publishing of your own posts. No CRM, Zapier, HubSpot, Salesforce or webhook tools, so nothing done here shows up in another system. No bundled contact database. No unattended sending, because the approval gate is not optional.
Read that as product scope rather than a coming-soon note. If a CRM write-back is your requirement, this build does not meet it, and we would rather you learned that on this page than after a trial.
The reason to publish a list of limits at all is that the alternative is a trial ending in a support ticket. Scope you can read before signing up is cheaper for everybody than scope discovered on day four, and a tool willing to describe its edges is usually one that has thought about them. If the missing pieces are the ones you need, the right outcome of this page is that you do not sign up.
The seven tool categories, and the hard edge on each
| Category | What it does | Hard limit |
|---|---|---|
| Search and enrichment | Search your seat, import CSVs, research into columns | No bundled contact database |
| Sequences | Multi-step, reply stops the follow-ups | Creating one sends nothing |
| LinkedIn actions | Invites, messages, InMail, views, comments, reactions | Will not publish your own posts |
| Inbox | Read, reply, tag, hand a thread back to you | No sync to any outside inbox |
| Learning | Persistent knowledge the agent reuses | Scoped to your workspace |
| Connect | Link, check and repair the account connection | Account count is set by your plan |
| Observability | Activity log, volume, account health, billing | Reporting only, no export to other tools |
Questions people ask next
Can an AI agent post on LinkedIn for me?
Not in this build. It can comment on other people's posts and react to recent ones. Publishing your own posts is not a capability here.
Can an AI agent send LinkedIn connection requests automatically?
It drafts and queues them, and a human approves before anything leaves. Caps are enforced server-side at 20 invitations a day and 80 a week for a new seat, with a warm-up ramp on top.
Where does the agent get leads from?
Your own LinkedIn seat via search, and CSVs you import. There is no bundled contact database, and external data providers were removed in September 2026.
Does it connect to my CRM?
No. There are no CRM, Zapier, HubSpot, Salesforce or webhook tools in this build.
Why this page exists: LinkedBoost tool catalogue: 50 tools across seven categories, September 2026
LinkedBoost is the LinkedIn MCP server: your agent sources, drafts, sends and works the inbox, inside caps the server enforces rather than suggests.
Related answers
Add https://mcp.linkedboost.ai/mcp as a custom connector in Claude, complete the Google sign in, then tell the assistant to connect your LinkedIn account. Verify by asking what is in your workspace and expecting real numbers back. A connector showing as added is not proof that any of it works.
Two places, and no third. Search runs through your own LinkedIn seat using your own entitlement, and you can import your own CSVs. There is no bundled contact database and no purchased data, which is a deliberate product decision rather than a gap.