Hands on LinkedIn
Search, signals, connection requests, messages, comments, reactions, posts, inbox - all as tools your agent calls in plain English, or endpoints you call from code.
For GTM engineers
The LinkedIn execution layer for GTM engineers
You already run your GTM motion as code. Lists as config, agents as workers, the CLAUDE.md as the constitution. The email leg is solved - cold email has a dozen reference architectures. LinkedIn is the channel everyone skips, because the only way to touch it was a cookie-stealing browser bot that gets accounts banned. Zevari is the hosted LinkedIn execution layer that closes the gap: 60+ tools your agent drives over MCP for Claude Code and Codex, or our REST API from your own code - signal-based targeting, Voice DNA, and approval gates on every write.
This page is for the person who builds the engine
If you run Claude Code or Codex daily, write your ICP as a config file, and think of outbound as a system to be versioned rather than a headcount to be hired - this is your page.
We merged the old AI engineer, GTM operator, and Claude / Codex power-user pages into one, because they were all describing the same person: the GTM engineer. You don't want a dashboard with 40 toggles. You want a tool surface your agent can reason over - over MCP for Claude Code and Codex, or our REST API from your own code - hosted state so the queue runs when your laptop is closed, and receipts on the safety claims. That's the whole pitch.
Not technical, or you're the founder who'd rather approve than operate? You're on the wrong page, and that's fine: we run it for you.
Connect
Add the server. One line. Or skip MCP and hit the API.
Reach Zevari over MCP for Claude Code and Codex, or our REST API from your own code. Hosted HTTP transport - nothing to install, no Docker, no Chrome profile to babysit.
MCP, one line
Works with Claude Code, Cowork, Codex, and any MCP client. The first tool call kicks off OAuth - you pick the org and LinkedIn account the session runs as.
claude mcp add zevari https://mcp.zevari.ai/mcpConfig as code
Drop it in your project .mcp.json and commit it - the whole team gets the same LinkedIn execution layer on clone.
{
"mcpServers": {
"zevari": {
"type": "http",
"url": "https://mcp.zevari.ai/mcp"
}
}
}Prefer the API? Most engineers do.
MCP is the fast path for Claude Code and Codex. But Zevari is reachable over MCP for Claude Code and Codex, or our REST API from your own code - and for many engineers the API beats MCP, because it drops straight into your existing pipelines, cron jobs, and workers with a bearer token. Same 60+ capabilities, same approval gates, same hosted state.
# The same capabilities over the REST API. Keys start with zvr_api_
curl https://api.zevari.ai/v1/contentTemplates/list \
-H "Authorization: Bearer $ZEVARI_API_KEY"
# Full OpenAPI reference: https://docs.zevari.ai/api/referenceFor the full onboarding - OAuth gotchas, billing, first three prompts - see the quickstart.
The gap
Why not just point raw Claude at LinkedIn?
Because Claude can find the needle in the haystack - and then it's handicapped. It can research a prospect beautifully and cannot connect to LinkedIn to act on it. And Claude's own scheduled routines hit a wall the moment they try to run outbound: they need human sign-off and have no persistent place to hold a multi-step campaign. That's the gap Zevari fills, and it's two specific things a bare MCP client or API caller can't do alone.
Hosted state
Campaigns, schedules, working hours, and approval queues live on our infrastructure, so the sequence advances on its own cadence whether or not your session is open. Persistent scheduling is the thing Claude Code structurally cannot do alone.
You keep the human-in-the-loop gate. We just make it a one-tap approval instead of a reason the whole thing never ships.
The surface area
60+ tools, grouped by what they do
You won't call these by hand - the agent selects them, or you call the matching REST endpoints. But here's the surface area so you know what you're holding.
Search + targeting
Profile, company, and post search; signal scanning; ICP scoring 1-5 with reasons; lookalikes from a single profile URL.
Research + voice
Person and company research, behavioral profiling, Voice DNA, communication blueprints.
Campaigns
Create multi-step sequences, add targets, advance steps, track progress, validate before send.
LinkedIn actions
Connection requests, messages, comments, reactions, profile views, posts - all approval-gated.
Inbox Radar
Classify replies by intent, enrich senders, stage drafts, send digests.
Content
Draft, schedule, and approval-gated publishing in your voice.
Pipeline
Propose meetings, opportunities, and closed deals into a lightweight CRM view.
Safety + settings
Live safety status, connection limits, working hours, pause and resume everything.
The full reference, with arguments and return shapes, lives in the MCP and API docs.
LinkedIn as code
The features, named - because they're invisible on most tools
Each one is a capability your agent reaches for without you micromanaging it.
Voice DNA
Trained on your sent messages, so drafts read like you wrote them, not like an AI wrote them. The same voice across every connection note and reply.
Signal-based targeting
Not a static title list. Find people who posted in the last 30 days about a topic, scrape commenters and likers of a viral post, then ICP-score every author 1 to 5 with the reason attached. Reach only the high-intent ones.
Campaigns + multi-step sequences
Your agent builds the sequence (warm action, connection note, value message), adds targets, and advances steps on schedule.
Inbox Radar
Every reply classified by intent (interested, objection, not-now, referral), senders enriched, drafts staged. You stop reading the inbox and start approving it.
Warm-by-default
Comments, reactions, and profile views before the ask, so the connection note does not land cold.
Hosted state
Persistent scheduling, working-hours queues, and approval gates that survive your session closing. The thing Claude Code or Codex cannot do alone.
Full capability tour on the product page.
A real session
End to end, in one chat
Actual run - prompt, staged campaign, chat approval, live. Only edit is trimming the target list for length.
You › Find people posting about "switching off Apollo" or
"Apollo alternative" in the last 30 days - B2B founders
and heads of growth. Score them against my ICP and show
me the top 5 with reasons.
Claude › [linkedin_search_posts] scanning last-30-day posts...
[agents_icp_score] scoring 41 authors 1-5...
Top 5 (score / reason):
5 Dana R., Head of Growth, devtools SaaS
- "Apollo's data decay is killing our reply rate."
Exact-fit ICP, active buying signal, decision-maker.
5 Marcus T., Founder, 12-person RevOps agency
- Asked for "Apollo alternative with better LinkedIn."
Agency = multi-account expansion.
4 Priya S., VP Sales, Series A fintech
4 Owen K., Founder, B2B data startup
3 Lena M., SDR manager
You › Build a warm-by-default campaign for the two 5s. Warm
action first, then a connection note that references their
post, then a value message. Use my voice. Stage it.
Claude › [agents_communication_blueprint] loading Voice DNA...
[campaigns_create] "Apollo-switchers - Jun" (2 targets, 3 steps)
[agents_generate_warming_comment]
[agents_generate_connection_note]
[agents_generate_messages]
[actions_request_confirmation]
Staged. To approve in this chat, reply exactly:
approve campaign Apollo-switchers
You › approve campaign Apollo-switchers
Claude › [actions_chat_approve_confirmation] approved
[campaigns_update_status] live
Live. Warm actions queue inside your working hours and
stay under your weekly connection ceiling. I'll classify
replies with Inbox Radar and stage drafts as they come in.Nothing touched LinkedIn until you typed the approval phrase. The drafts came from Voice DNA on your sent-message style. And because the queue is hosted state, it keeps advancing on schedule with your session closed.
Outbound as code
Bring your own CLAUDE.md
Run Zevari the way you'd run any serious agent - a constitution and your ICP as config, version-controlled in the repo. We publish the reference template so you're not starting from a blank file: a LinkedIn CLAUDE.md constitution (tone rules, approval policy, do-not-contact lists) plus an icp.yaml your agent scores against.
Every existing reference architecture for this is cold-email-first. This is the LinkedIn one. Build on it, fork it, PR it.
outbound-as-code/
├── CLAUDE.md # constitution: voice, approval policy, DNC, working hours
├── icp.yaml # who counts as a 5, who counts as a 1, and why
├── campaigns/ # versioned sequence definitions
└── routines/ # scheduled jobs: morning digest, daily connect, inbox triageSafety, in receipts - not adjectives
You're going to ask "is this going to get my account banned?" before you ask anything else. Here's the mechanism, not the marketing.
- Every write - message, connection, comment, post - is staged for human approval before it sends.
- Session-based connection, no browser cookies, no password handoff, no extension.
- Weekly connection ceilings: Free 40, Premium 80, Sales Navigator 150.
- Working hours, behavioral pacing, duplicate checks, and burst caps - enforced server-side, not suggested.
- A year of refinement, zero ban incidents.
The math you already did
You priced the alternatives before you got here
OptionCostThe catch
In-house SDR$110-160k/yr fully loadedYou manage a human
AI SDR (11x, Artisan)$45-60k/yr50-70% churn; one got LinkedIn-banned
GTM engineer salary (median)$127k/yrThat's you
LinkedIn agency$3-5k/moA black box you can't inspect
Zevari Operator$87/moYou run it, and you approve every send
Additional LinkedIn accounts are $37 each - versus the $79-99/account everyone else charges. Scale with systems, not headcount.
FAQ
GTM engineer FAQ
Do I run Zevari in Claude Chat, Cowork, or Claude Code?
Any of them, plus Codex and any other MCP client. The hosted HTTP server works wherever MCP works, and if you'd rather skip MCP entirely you can call our REST API straight from your own code. Most GTM engineers run it in Claude Code or Codex so the CLAUDE.md constitution and ICP config live alongside the rest of their stack.
Can I use the REST API instead of MCP?
Yes - this is the engineer's preferred path. Reach Zevari over MCP for Claude Code and Codex, or hit our REST API from your own code: same 60+ capabilities, same approval gates, same hosted state. For many engineers the API beats MCP because it drops straight into existing pipelines, cron jobs, and workers with a bearer token. Full reference at docs.zevari.ai/api/reference.
Why can't I just use Claude's own routines for LinkedIn outbound?
Two reasons. Claude can't connect to LinkedIn to act - the API is too restrictive for an unaided agent. And routines need human sign-off with no persistent place to hold a multi-step campaign, so the sequence never actually runs. Zevari gives the agent hands on LinkedIn plus hosted state for the schedule and approval queue, while keeping the human-in-the-loop gate as a one-tap approval.
What do I need open on my computer to use it?
Just an MCP client connected to Zevari, or nothing at all if you call the REST API from a server. No browser extension, no Chrome profile, nothing running in the background. State is hosted, so scheduled actions run on our infrastructure even when your session is closed.
Can it find people who are posting about a topic and reach out to them?
Yes - that's signal-based targeting. Scan posts for intent keywords from the last 30 days, scrape commenters and likers of a specific post, ICP-score every author 1 to 5 with reasons, and have the agent reach only the high-intent ones.
How does the ICP scoring decide a 3 vs a 5?
It scores against the ICP you define (the icp.yaml in the reference template, or in-product config), returns a 1-5 with a written reason per prospect, and lets you correct it. If a score looks wrong, you tell the agent why someone isn't your ICP and it adjusts - the reasons are visible so you're never trusting a black-box number.
Will this get my LinkedIn account banned?
No incidents in a year of refinement. Connection is session-based with no browser cookies, every write is approval-gated, and weekly connection ceilings, working hours, behavioral pacing, and burst caps are enforced server-side. Full mechanics on the safety model page.
Do I need LinkedIn Sales Navigator?
No. You can start on a free account. Sales Navigator only raises your weekly connection limit (Free 40, Premium 80, Sales Navigator 150).
Can I run outbound for multiple clients or companies?
Yes. Each LinkedIn account connects in its own workspace, and additional accounts are $37/mo - versus the $79-99/account most tools charge. Agencies run a workspace per client. For full done-for-you multi-account operation, see the managed page.
Does it actually send the email leg, or just draft it?
Zevari runs the LinkedIn leg end to end. For email follow-up to non-responders, the agent enriches a verified email from the LinkedIn profile and hands it to your sender of choice (Instantly, Smartlead, Resend). You approve before anything sends.
Connect to Claude Code
Operator - $87/mo. Additional accounts $37 vs the $79-99/account industry standard. Reach Zevari over MCP for Claude Code and Codex, or our REST API from your own code. Free to try, cancel anytime.
claude mcp add zevari https://mcp.zevari.ai/mcpWant it installed with you? Launch - $1,500: we pair to wire the constitution, your ICP config, and your first live campaign together.