AI connectors: Claude, Claude Cowork and other AI apps
Your company decides whether its people may connect AI apps such as Claude and Claude Cowork to TRANXACT. An administrator accepts a statement for the company, each person approves each app by name, every app acts only as that person, and everything is in the audit log.
Since release 20261106g, people can connect AI apps they use (Claude, Claude Cowork, Claude Code and other apps that speak the Model Context Protocol, MCP) to TRANXACT and ask them about their own data. TRANXACT provides the framework: the sign-in, the controls and the audit log. Whether it is allowed is your company’s decision, not TRANXACT’s.
Three decisions, three owners
- Your company decides whether AI apps may connect at all. A company administrator, signed in as themselves, reads a statement, types their name and title and accepts it for the company. Until then, no AI app can connect.
- Your company decides what AI apps may reach. Nothing is reachable by default: the company publishes data to the AI-CONNECTORS client in Platform › API Studio, as it would for a partner’s API key, and chooses whether connectors may only read or may also write.
- Each person decides whether their AI app may act for them. They sign in to TRANXACT, see the app’s name and what it could reach, and approve or decline it. The app then acts only as that person, so their own roles and business units still apply: it never sees more than they could see on screen.
Allowing AI connectors for the company (administrators)
- Open Platform › API Studio and go to AI Connectors.
- Choose Review and allow AI connectors and read the statement.
- Type your full name and title. Optionally, list the only places apps may return to after a person approves them (for example claude.ai), and tick whether connectors may also create, change or delete records where you grant it. Without that tick they only read.
- Tick I accept this statement on behalf of the company and choose Allow AI connectors.
- Publish what connectors may reach: grant exposures to the AI-CONNECTORS client, exactly as for any API client.
- Give your people the connector address shown on the panel. It is your TRANXACT address followed by /api/v1/mcp.
This must be done by the company’s own administrator, signed in as themselves. It is refused from a session held on someone else’s behalf (a TRANXACT operator or support person using the administrator’s name), from an accountant’s seat, from a service account and from ASK. The permission is the company’s.
Connecting your AI app (everyone)
- In your AI app, add a custom connector (in Claude: Settings › Connectors › Add custom connector) and paste your company’s connector address.
- The app opens TRANXACT. Sign in as you normally do.
- Read what the app is asking for: its name, your company, where it returns to, and what it would be able to read or change as you. Choose Approve, or Decline.
- Ask the app about your data. It uses the tools your company published, as you.
Preferences › AI connections lists the apps you approved, when, and when each was last used, with Disconnect beside each. Disconnecting stops the app at once.
Withdrawing, and disconnecting someone
- An administrator can choose Withdraw for the company, with a reason. Every person’s connection ends at once and no app can connect until the company allows it again.
- An administrator can disconnect any one person’s app from the Connections list on the AI Connectors panel.
- The company’s decisions are kept, with who, title, the statement accepted word for word, when and from which address, and who withdrew it and why.
What is recorded
- In the Audit Log (mandatory, cannot be switched off): the company accepting and withdrawing, each person approving, declining and disconnecting an app.
- In the API Call Log: every call an AI app makes, with the person and the app.
- On each record an AI app creates or changes: the change is the person’s, labelled with the app (for example ai:Claude), so it can be told apart from the same change made on screen.
보안
- Sign-in uses OAuth 2.1 with PKCE. TRANXACT never sees the app’s password, and the app never sees yours.
- An app gets a token that lasts an hour and is renewed while the connection stands. Tokens are stored only as fingerprints, and a renewal token used twice ends the connection.
- A connector token opens only the MCP endpoint, not the rest of the API and not the console.
- What an AI app reads is sent to that app’s provider, under your company’s own agreement with them.
Any AI platform
- Claude, Claude Cowork and Claude Code: add a custom connector with your company's connector address; Claude registers itself and opens TRANXACT for you to approve it.
- ChatGPT: in developer mode, create a connector with the same address and OAuth. If your company limited the places apps may return to, the administrator adds chatgpt.com.
- OpenAI Responses API and Agents SDK, Gemini, LangChain Deep Agents and other MCP clients: connect with OAuth if they can register themselves, or have an administrator register them.
- Microsoft Copilot Studio and other platforms that cannot register themselves: an administrator chooses Register an app or a headless agent on the AI Connectors panel, picks An app people sign in to, and pastes the client id, secret and addresses shown (once) into the platform.
Headless agents: harnesses and loops on a server
An agent that runs on a server without a person at a browser (a harness built with an agent SDK, or a loop that runs every night) is registered by an administrator as A headless agent, with what it does and the person it runs as. It does nothing until that person approves it under Preferences › AI connections › Agents that run as you. It then acts only with that person's permissions, is listed on the AI Connectors panel, and can be retired there, which stops it at once.
In Agentic AI, beside TRANXACT's own agents
- Agentic AI › Agents (Agent Studio): every connected app or headless agent is an agent with Origin external, owned by and running as its person. TRANXACT never runs it: its model and loop are on its own platform.
- Agent Runs: each day of work through a connection is a run, closed when the connection ends.
- Tool Calls: every call it made, with the arguments, whether it was allowed or refused, how long it took and how many records.
- Delegated Tasks, ASK hand-offs and TRANXACT's own agents are unchanged and appear in the same registers.
For developers: the server publishes OAuth metadata at /.well-known/oauth-protected-resource and /.well-known/oauth-authorization-server, supports dynamic client registration, and answers an MCP call without a token with 401 and a resource_metadata pointer, as the MCP authorization specification describes. See API endpoint reference: how to call TRANXACT.