# AI agents How to connect Claude, ChatGPT, Claude Code or any MCP agent to your workspace, what an agent is allowed to do, and how to disconnect one. AI agents is where you connect your own AI to PYLO. Once connected, Claude, ChatGPT or any agent that speaks MCP can answer questions about your workspace and set things up for you to approve, such as onboarding a new client. ## Who can open it Everyone in the workspace. Each person connects their own agent, and the agent works with that person's role and never more. Owners, admins and ops users see every agent connected to the workspace. Everyone else sees the agents they connected. ## The server address The card at the top shows PYLO's MCP server address, `https://api.pylosystem.com/mcp`, with a `Copy` button. It is the same address for every agent. When the agent signs in, PYLO works out which workspace and which person it is acting for. ## What every agent is held to - It acts as the person who approved it, and never above their role. - It never sees your password or any connected tool's credentials, and it cannot move money. - Configuration changes arrive as a plan you approve under [Plans](/docs/settings/plans) before anything changes. - You can disconnect it at any time, and it stops working on its next request. ## What you let it do When you sign an agent in, PYLO lists everything your role can allow, ticked, whatever the agent asked for. Leave them ticked and the agent can onboard clients, set up workspaces, create connect links and send test leads, with every change still arriving as a plan for you to approve. Untick anything you want it not to do. Managing tokens starts unticked. On both plans an agent can read everything its person may see. Planning, applying and undoing changes, onboarding a client, sequences and provisioning need the [Pro plan](/docs/billing-and-plans): on Basic the agent is refused and told the plan it needs. If an agent later needs something you did not allow, it can ask you for it. It gives you a link to PYLO's own screen, which shows what the agent can do now, what it is asking for and its reason. Tick what you allow and the agent can use it on its next step, with the same connection. Only you can allow it there, never the agent. Connecting the same agent again works too, and replaces its old sign-in rather than keeping both. ## Connect Claude 1. In Claude, open Settings, then Connectors, and choose Add custom connector. 2. Name it PYLO and paste the server address. Leave the advanced settings empty. 3. Choose Connect. PYLO asks you to sign in, pick the workspace or your agency, and choose what Claude may do. 4. In a chat, switch PYLO on from the tools menu. On a Claude Team or Enterprise plan, an owner of your Claude organisation adds the connector once. Each person then connects with their own PYLO sign-in. ## Connect Claude for your whole agency If you are in an agency, you can connect Claude once for every client instead of one workspace at a time. Add the connector as above, and when PYLO asks what Claude may act for, choose "All of" followed by your agency's name. Claude can then work in every client workspace and, on Pro, set up new ones from a chat: ask it to onboard a client, and it asks you the setup questions, drafts the new workspace as a plan, and waits for your yes before anything is created. The same card is in Agency OS, under Settings. ## Connect ChatGPT 1. In ChatGPT, open Settings, then Apps, and turn on developer mode under Advanced settings. 2. Choose Create. Name it PYLO and paste the server address. 3. Choose OAuth for authentication. Leave the client fields empty, because PYLO registers ChatGPT itself. 4. Choose Create, then approve on PYLO's sign-in screen. ChatGPT Business, Enterprise and Edu can use every action you allow. On ChatGPT's own Pro plan, ChatGPT limits apps like this one to reading. ## Connect Claude Code Run `claude mcp add --transport http pylo https://api.pylosystem.com/mcp` once. Then, in Claude Code, type `/mcp`, choose pylo and choose Authenticate. PYLO's sign-in opens in your browser. ## Other agents Cursor, Windsurf, n8n or your own agent can connect with the same address. If the agent supports OAuth sign-in, you approve it the same way as Claude. If it only accepts a header, an owner, admin or ops user creates a token under [API tokens](/docs/settings/api-tokens), and the agent sends it as `Authorization: Bearer`. ## Connected agents Each connected agent shows its name, when it was connected, when it was last used, and what it may do. Owners, admins and ops users also see who connected it. An agent connected for your whole agency is marked "for your agency". Choose `Disconnect` to stop an agent. PYLO asks you to confirm first. The agent stops working on its next request, and connecting it again needs a new sign-in. Every tool an agent can call is listed in the [MCP server reference](/docs/mcp). ## Unused connections are disconnected A connection nobody has used for 30 days is disconnected automatically by a daily check, and so is an API token. Each connection on this page shows its permissions, when it was last used, and the date it will be disconnected if it is not used before then. Connect the agent again when you need it.