Reachkit MCP¶

The outreach pipeline that most operators run across four tools now sits in one conversation: find the people, import the emails, write the sequence, launch it, then answer what comes back.

Server type: Remote (Streamable HTTP)
Endpoint:    https://reachkit.ai/mcp
Transport:   JSON-RPC over POST
Auth:        OAuth 2.1 with dynamic client registration, or an account access token
Scopes:      mcp (plus openid, email for sign-in)
Tools:       lead search, contact import, list research, sequence builder,
             campaign launch and analytics, Unibox replies
Built by:    Reachkit (reachkit.ai)
Registry:    ai.reachkit/reachkit version 1.0.0, active

Why This Matters for Operators¶

Cold email is the one acquisition channel where the work is entirely mechanical and entirely sequential: search a list, enrich it, write the sequence, attach an inbox, schedule it, then read replies and answer them. That is six tools and a spreadsheet in most companies, and the handoff between each one is where campaigns stall.

Putting it in the assistant does not change the strategy. It changes how much of the strategy actually gets executed, and it is the first server in this catalog where an agent can spend sending reputation rather than money.

What It Actually Does¶

initialize without a token returned HTTP 401 with {"error":"unauthorized"}. The discovery documents answer without an account:

/.well-known/oauth-protected-resource returned HTTP 200: resource https://reachkit.ai/mcp, authorization server https://reachkit.ai, bearer_methods_supported: ["header"], scopes_supported: ["mcp"].
/.well-known/oauth-authorization-server returned HTTP 200 with a registration endpoint at /oauth/register (RFC 7591, so a client registers itself), token_endpoint_auth_methods_supported: ["none"] (public client), PKCE S256 only, authorization code and refresh grants, authorization_response_iss_parameter_supported: true, a userinfo endpoint, and identity claims sub, email, email_verified.

The vendor documents the surface as six flows:

Flow	What the agent does
Find leads	Search 500M+ people and 70M+ companies by title, seniority, industry, location, company size and technology. Searching is free inside a daily allowance.
Import contacts	Import with verified work emails at one credit each, and matches without a verified email are refunded.
Research the list	Break a list down row by row and save findings as columns that become merge tags.
Build campaigns	Write the sequence with A/B variants, merge tags and spintax, and preview every email as it will send.
Launch and manage	Attach inboxes, set the schedule, pause and resume, read analytics by campaign and reply rate.
Answer replies	List, label and answer Unibox conversations, on the Scale plan.

The Part That Costs You Something¶

Two things here are not reversible in the way a document edit is.

Launching spends sending reputation. A campaign goes out from inboxes that are attached to your domains, and the blocklist and bounce handling that come with the account are what stand between a bad import and a burned domain. The vendor states that the assistant shows the campaign and waits for a go before launching, and that drafts and previews never send - but that gate lives in the client's approval behaviour, not in the token scope, because the scope is a single mcp.

Imports spend credits. Verified-email imports bill one credit each, with a refund when no verified email is found. The search itself is free within the daily allowance, so the cost is in the import, which is the step worth pricing before it runs.

How It Is Contained¶

OAuth sign-in as you. The assistant acts with your workspace, your member permissions and your plan limits, and access is revoked from the account page immediately.
Identity scopes only. mcp, openid and email - no per-object permission split, so the containment is the account role plus the approval step.
Previews are free. Sequence previews and drafts do not send, which makes the review step genuinely non-destructive.
Private client registration. token_endpoint_auth_methods_supported: ["none"] with mandatory PKCE is the correct shape for a desktop or CLI client that cannot keep a secret.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations Recorded¶

No tool call was executed: every path requires an account, so the capability groups come from the vendor's integration page rather than a live tools/list.
Reply handling is plan-gated. Reading and answering the Unibox requires the Scale plan; a lower tier gets find, import, sequence and launch only.
One scope for everything. An agent that can read analytics can also launch a campaign, so the operator's approval behaviour is the only brake on the spending path.
Lead counts are directory matches, not intent. Title and technology filters select who exists on a list, not who is buying this quarter.
Importing contacts a second time is a paid import again unless deduplication against existing campaigns is done first, which the vendor documents as part of the import step.

FAQ¶

Is this a scraping tool for LinkedIn?¶

No. It searches Reachkit's own people and company index, which is why searching is free and importing verified emails is what costs credits.

Can the agent send email without me seeing it?¶

Launching is documented as an approval step, and previews never send. The token carries one scope, so the protection is the client's confirmation, and it is worth testing that your client asks before you attach a real sending inbox.

What should I actually test first?¶

One narrow segment and one sequence. The expensive mistake is not the credit spend, it is a 200-contact send from a domain whose reputation you need for the next year.