AI employees on your payroll, not someone else's
Most software is bolting on a vendor's chatbot and charging you for the privilege. Coordinator goes the other way: hire AI agents as real team members with real permissions, running on your API key — and connect the AI you already use, like Claude or ChatGPT, to run your whole business conversationally.
Free. No card. About two minutes to set up.
Agents that hold real roles in your team, run on your own API key, and work your books, projects and records all day — with a kill switch.
How employees workChatGPT, Claude, Gemini, Grok, Codex or Antigravity runs Coordinator through our built-in MCP server — on a scoped, expiring token that can't send, pay or delete.
The 5-minute setup guideAn agent is a team member, not a feature
In Coordinator, an AI agent joins your organization the way a person does: it holds roles, gets exactly the permissions those roles grant, and shows up in the same team settings — no parallel 'AI permissions' system to misconfigure. Which means its ceiling is its roles' ceiling: whatever a person with those roles could do here — across the books, the projects, the schedule, the records, the threads — you can delegate.
- Real roles, real fences — give an agent the same role bundles as staff — per-module permissions, entity ring-fencing, the works. It sees only what its roles reach.
- Plugged into your workflows — agents work the same objects your team does — triaging intake, drafting documents, preparing invoices, organizing records — inside the processes you already run.
- Your API key, your model bill — agents run on your own provider key, encrypted like every other credential. No markup, no metering, no lock-in to a model you didn't choose.
- Everything opens closed — a new agent can do nothing: no instructions, no delete rights, no outbound, empty reach — until you explicitly grant each one.
- Pause is a kill switch — pausing an agent is immediate and total. And an agent can never hold the Owner role, by rule.
- Attribution is mandatory — everything an agent does is labeled as agent work in the audit trail — there is no setting to make AI look human.
The dangerous part of agent AI is reach. Coordinator makes reach a ladder you climb deliberately — and the counterparty gets a say.
The default. Your agent drafts, organizes and prepares; a human sends. It literally cannot message anyone outside the org.
The agent may contact a connected business only if that business has opted in to receiving agent messages. Their inbox, their rule.
Full reach, still attributed, still rate-limited (5/recipient/hour, 60/agent/hour), still failing closed. You grant this; it's never assumed.
And on the receiving side: your organization decides whether other companies' agents may message you at all — allow, draft-for-review, or never. Default: draft-for-review, because nobody consents by not having heard of a setting.
Their agent works for them. Yours works for you.
Claude today, something better next year — your agents ride your provider account, so you upgrade when the field moves, not when a vendor renegotiates.
Your data is read under your roles, your key, your audit trail. A vendor's bundled AI is a second company inside your books; this isn't.
Permissions, reach, outbound level, pause — controls designed like employment, because that's what delegation is.
Already shipping in the product: forward a vendor PDF to your bills inbox and a vision model — on your own key — reads it into a draft bill, validated before it touches your books.
Bring a key from any of these
Each one connects in about a minute and gets the full treatment: its own stored key, a model list fetched live from the vendor, and a real validation call before anything is saved. Pick a brain per agent — and change your mind later without rebuilding the team.
The strongest document reader — and the natural pair for our MCP server.
Orchestrate agents on ClaudeNative PDF reading and Google-scale context windows on your own key.
Orchestrate agents on GeminiGPT models — plus a base-URL override that reaches any OpenAI-compatible endpoint.
Orchestrate agents on OpenAIOne key that reaches Claude, Gemini, GPT and the open-weight field.
Orchestrate agents on OpenRouterPrivate inference — open weights first, frontier models routed, prompts not stored.
Orchestrate agents on VeniceMax-efficiency inference for long-running agent work — their claim: up to 10x cheaper per token.
Orchestrate agents on SailThe Meta Model API's Muse Spark models — with the contributor-tier trade-off explained.
Orchestrate agents on MetaxAI's Grok models, with strict two-step key validation.
Orchestrate agents on GrokSettings → AI: one row per provider, each with its own stored key, live model catalogue and connect-time validation.
The OpenAI row also accepts a custom base URL, so any OpenAI-compatible endpoint — an Azure-style gateway, LiteLLM, a local proxy, an on-prem model server — works too.
Run Coordinator from Claude, ChatGPT — or any AI you like
Coordinator ships a built-in MCP (Model Context Protocol) server. Mint a connector token in settings, paste it into Claude, ChatGPT or any MCP-capable client, and your AI can work your actual business: look up a client's history, create the follow-up tasks from a meeting, log your time, stage a draft invoice.
- One token, standard protocol — works with any MCP client, today's and tomorrow's. No zapier-glue, no custom integration project.
- Scoped like a person, not a god-key — the token gets the intersection of what you granted and what your live role allows, re-checked on every call. Revoke it, or lose the role, and it dies instantly.
- Reversible by design — your AI can read, create work, log time, add contacts and stage draft invoices. It structurally cannot send an invoice, move money, message a counterparty, approve anything or delete anything. Ever.
- Audited under your name — every connector action lands in the audit trail, marked as connector work, expiring in 90 days unless you renew it.
"What's outstanding with Harbor & Finch?" → your AI reads the contact, the open invoices, the project status. "Draft the final invoice and make follow-up tasks from yesterday's meeting notes" → a draft invoice and owned tasks appear — for a human to review and send. The AI does the legwork; the send button stays yours.
Put your AI to work
Free like everything else — bring your own key, mint a connector token, and let your AI do the busywork inside real guardrails.
Free. No card. About two minutes to set up.