Which AI agents can connect?

Any agent that speaks MCP. There are step-by-step guides for ChatGPT, Claude, Grok, Gemini and others on the agents page, each read off that agent’s own documentation.

Can my agent send emails or publish posts?

Not on its own. Sending and publishing run from the queue after an approval exists, and approving is a separate, explicit call that only a credential you granted that permission can make. Drafts an agent writes land in the same queue as yours.

Agent access comes with the paid plans, from Starter at $99 a month; the Free plan does not include it. Each tool says in its description whether it is cheap or expensive, so an agent can batch instead of looping.

How do I take access away?

Revoke the connection or the key in Settings, Agent access. It stops working on the next call.

Your GTM team, inside the agent you use

Ask your agent what needs you today, which companies changed, or to draft a post. Every draft still waits for your approval.

Get my free GTM plan See pricing

OAuth or a scoped key
Only the ventures you name
Approvals still apply

The short version

InstinctGTM runs an MCP server, so Claude, ChatGPT and any other agent that speaks MCP can read your brief, search the companies and people you track, capture what it finds and draft posts and messages into your approval queue. It runs under a credential you issue, reaching only the ventures you name, and every gate, approval and limit the screen has applies to it too.

Read against the registry on 2026-10-13

Set it up

1
Copy the address
Open Settings, then Agent access, and copy the MCP address.
2
Add it to your agent
Paste it where your agent adds a connector or MCP server. It sends you back to choose what it may do and which ventures it may reach.
3
Ask
Ask in plain words. The agent picks the tool; you see what it called in the call log, including anything refused.

Setup guides for ChatGPT, Claude, Grok and Gemini

What you can ask

“What needs my approval today?”
get_brief
“Which drafts are waiting for me?”
list_approvals
“Who do I owe a follow-up?”
list_followups
“What changed at the companies I watch?”
list_companies
“Add this company to my watchlist.”
watch_company
“Draft a LinkedIn post from this idea.”
draft_post

Every tool it exposes

30 tools, read from the registry the server runs on.

List venturesReads only
Today's briefReads only
What is waiting on approvalReads only
Overdue follow-upsReads only
Find companiesReads only
One company, with its timelineReads only
Find peopleReads only
One person, with their historyReads only
Find contentReads only
List scheduled and published postsReads only
Channel performanceReads only
What the system learnedReads only
Mentions and capturesReads only
Watch a companyChanges something
Add a personChanges something
Log a touchChanges something
Capture a threadChanges something
Submit a content ideaChanges something
Read a company nowChanges something
Draft a post for the calendarChanges something
Draft content for every channelChanges something
Draft a reply to someone who engagedChanges something
Draft the next message in a sequenceChanges something
Approve one piece of contentChanges something
Schedule a postReaches somebody outside
Approve one outreach draft, and send itReaches somebody outside
Mark a post as publishedChanges something
Move a relationship to a new stageChanges something
Push a follow-up outChanges something
Why the system concluded thatReads only

What no credential can do

send or publish
Sending and publishing go out under the founder’s name. They run from the queue after an approval row exists, and creating that row is a separate, explicit call. An agent that can approve and send in one step is an agent that can publish without anybody reading it.
raise a limit, a cap or an autonomy stage
Deliverability caps, per-domain limits and the autonomy ladder are the blast-radius controls. A caller that can widen its own limits has no limits.
read across ventures without a grant
A venture is the isolation context. A key names its ventures and every read is filtered by that list, not by what the caller asks for.
bulk export contacts or message bodies
Page-at-a-time reads serve every legitimate agent workflow. An unbounded dump of people and message bodies is the shape of an exfiltration, and it is not needed for anything the product does.
write to the learning boundary
Outcomes are recorded by the pipeline that observed them. A caller that can write learning rows can teach the allocator anything it likes, and the allocator moves real budget.
reach billing, keys, or another member’s session
Account administration is a human surface. Nothing about an agentic workflow requires it, and everything about it is worth a password prompt.

Questions people ask