POSTARD DEVELOPERS
MCP integration for AI tools
Give your AI client a connection to your content workspace, with explicit permission for what it can see and save.
How the connection works
- A compatible AI client connects to the configured API origin at
/mcpusing Streamable HTTP. - OAuth sends you to Postard to sign in and select workspaces. Saving drafts is a separate opt-in permission.
- The client receives a token bound to the registered MCP resource. Every tool call checks current consent and workspace access.
No API secret belongs in the conversation. OAuth is the supported authorization flow. Generic app-session tokens cannot be used at this endpoint.
The five available tools
| Tool | Input | Behavior |
|---|---|---|
get_workspaces | None | List only consented workspaces. |
get_drafts | workspace_id | Read up to 50 recent drafts. |
create_draft | workspace_id, title, content, request_id | Save a text draft with separate write consent and an editing role. |
get_connected_accounts | workspace_id | Read profile metadata and connection status. |
get_account_performance | workspace_id, connection_id | Read a sourced, dated metric snapshot, or an explicit unavailable result. |
IDs are UUIDs. Draft titles are 1–120 characters; content is 1–10,000. Successful tool responses include text content and structuredContent.result. Tool failures set isError: true with a safe message. None of these tools publishes or schedules a post. Workspace-scoped calls require an active plan or trial. For create_draft, generate one request_id UUID and reuse it only for retries of the same content.
{
"workspace_id": "11111111-1111-4111-8111-111111111111",
"request_id": "22222222-2222-4222-8222-222222222222",
"title": "Friday reflection",
"content": "One thing we learned this week…"
}Client setup for a configured instance
Hosted use requires an HTTPS API, Supabase OAuth server, and a registered client. This release does not offer automatic public registration or a one-click marketplace listing.
- The instance operator registers an OAuth client with the exact callback URL required by the AI client. Use PKCE and configure the appropriate public or confidential client type.
- The operator allows the client ID in
MCP_CLIENT_IDSon both web and API, enables the access-token hook, and binds the registered client to the exact API/mcpresource. - Configure the AI client's MCP connection with that endpoint and registered OAuth client. Protected-resource metadata is published at
/.well-known/oauth-protected-resourceand/.well-known/oauth-protected-resource/mcp. - Complete Postard consent, then try
get_workspacesbefore using a workspace-specific tool. Check with the instance operator if setup is pending.
Claude, ChatGPT, and Cursor setup
These clients document remote MCP and OAuth. Postard uses registered OAuth clients; choose the static or own-client option in the client where available. Our hosted connection to each client still needs validation. An icon identifies a setup guide, not a marketplace listing or endorsement.
Claude supports custom remote connectors and an own OAuth client option. ChatGPT supports custom MCP servers with static OAuth credentials. Cursor supports static OAuth configuration in mcp.json. Follow the linked official guide for the current interface, callback URL, and account restrictions.
MCP setup guides. A configured Postard endpoint and registered OAuth client are required; hosted connections still need validation.
Bring Postard into Claude.
Copy this prompt into your AI. It can help configure the connection or walk you through the settings. You finish sign-in and choose workspace access.
This uses the URL you supplied; it does not verify that the server is connected.
Read the setup prompt
The prompt guides setup and can help a coding agent propose a configuration change. It cannot bypass client settings, OAuth consent, missing registration, or an unavailable hosted endpoint. No passwords, secrets, or access tokens belong in the prompt.
You stay in control
Read access is limited to selected workspaces. Draft-write consent is separate and does not elevate your role. Grants expire after 30 days; you can revoke them in workspace Settings. Revocation is checked on later calls. Social credentials are never exposed to the AI client.
Public ChatGPT distribution has not been submitted or approved. Compatibility with a specific hosted AI client must be verified during its setup. Read the current request limits and MCP allowances.