Procedure Platformby MethodMinded
Settings
← Resources
Connected assistance

Connect your AI through MCP

Add your installation’s server URL in an approved AI client, sign in, and choose the organization and access to allow. Procedure Platform keeps one set of tools and permissions across providers and deployments.

Connect to a cloud installation

  1. An organization administrator opens Settings → AI connections, enables access, and chooses eligible users, groups and capabilities.
  2. Copy the MCP server URL shown there. It is your installation’s public origin followed by /mcp.
  3. Add that URL in your approved AI client’s connector or custom MCP settings. Choose automatic OAuth authentication where available.
  4. Click Authenticate or Connect. Sign in to Procedure Platform, choose the organization, select the permitted capabilities, and authorize the connection.
  5. Ask the assistant to list your procedures or create a new procedure draft. Follow the returned link to review the result here.

Clients supporting Client ID Metadata Documents (CIMD) or Dynamic Client Registration (DCR) discover and identify themselves automatically. You do not create a client ID, secret or callback URL for these flows. Consent starts with draft writing off; select it only when you intend to allow new drafts and your organization permits it. Existing manual registrations remain supported in Advanced compatibility.

ChatGPT

  1. Open ChatGPT’s Plugins area and choose Add custom MCP server. Some accounts expose this through Apps or workspace developer settings.
  2. Enter your installation’s MCP server URL. Choose OAuth / Auto authentication and CIMD when offered; DCR is the compatibility option. Leave manual credentials empty for automatic setup.
  3. Create or install the connection as permitted by your workspace, authenticate, and complete Procedure Platform’s organization and consent screen.
  4. Select the installed application in your conversation and ask it to list procedures. To submit drafts, select draft-writing access during consent and obtain administrator write approval if your organization requires it.

If the desktop callback says “Authentication complete” while the app still offers Set up authentication, the browser has returned but the token exchange or first MCP request may have failed. Start a fresh Authenticate attempt after updating the installation; an earlier failed authorization code cannot be reused. Keep the exact callback address supplied by the client.

Custom MCP availability, write confirmations and workspace approvals depend on your ChatGPT account. A private installation additionally needs supported network connectivity. See OpenAI’s current setup instructions.

Claude

  1. In the connector settings available to your Claude account, add a custom remote MCP connector using your installation’s server URL.
  2. Use the automatic OAuth flow. Claude can use a published CIMD identity with a public-client token exchange, or DCR when needed.
  3. Authenticate to Procedure Platform, choose the organization and capabilities, and connect.
  4. Start with a procedure search or the authoring-format resource. Review submitted drafts in Procedure Platform.

Claude’s hosted connector and Claude Code use different callbacks. Automatic identification supplies these details; the advanced manual option must use the exact callback specified by the client. Claude Code’s native loopback callback may change port while retaining the registered host and path. See Anthropic’s authentication requirements.

Microsoft Copilot

Copilot Studio

  1. An agent maker opens the agent’s Tools page, then Add a tool → New tool → Model Context Protocol.
  2. Enter the Procedure Platform server URL and choose OAuth 2.0. Select Dynamic discovery when installation DCR is enabled.
  3. Create an end-user connection, authenticate to Procedure Platform, choose an authorized organization, and select the tools the agent may use.
  4. Have your Microsoft administrator review environment permissions, Power Platform data policies, connector approval and publication before making the agent available.

If enterprise policy requires a fixed client, have an administrator register an Organization application in Procedure Platform’s Advanced compatibility section, then enter its one-time credentials in Copilot Studio’s manual OAuth configuration. The issuer, authorization and token URLs are published by your installation’s discovery metadata. Do not configure a shared maker credential as if it identifies each end user. Microsoft’s onboarding guide.

Microsoft 365 Copilot agents

Use Microsoft 365 Agents Toolkit to create a declarative agent with an MCP plugin. Select dynamic OAuth registration where supported, or provision a fixed registration through Microsoft’s OAuth vault process. Toolkit generates the tenant-specific manifest and package. Provisioning, sideloading, admin approval and deployment belong to the customer’s Microsoft tenant; a generic server URL does not automatically install an agent.

Entra SSO and delegated Procedure Platform OAuth are different authentication choices. An Entra token does not grant Procedure Platform permissions automatically. This release supports the Procedure Platform OAuth route; custom Entra federation and SAML are not advertised as completed integrations. Microsoft 365 MCP plugin guidance.

Local clients and restricted networks

A local AI client can connect directly to the installation’s Streamable HTTP endpoint. For a client that only accepts stdio, the installation includes scripts/mcp-stdio-bridge.mjs. Run it with Node 20.19 or later and the installation’s MCP URL; configure the local client’s command to launch that script. It forwards the same tools and resources without adding a cloud relay.

The bridge uses local-installation OAuth with a loopback browser callback. Open the sign-in URL shown in the client’s bridge logs and complete consent. Tokens stay in memory; restarting requires sign-in again. Refresh works while the bridge runs. Administrators may alternatively inject a short-lived access token through PROCEDURE_PLATFORM_MCP_ACCESS_TOKEN; never place it in the server URL or command arguments.

An air-gapped installation can use local accounts, a compatible local model, and a manually registered local OAuth client when DCR is disabled. Disable CIMD if external metadata retrieval is prohibited. No MethodMinded cloud authentication, model, telemetry or transport service is required for MCP execution. Use a trusted internal TLS certificate and the local runtime’s CA configuration; do not bypass certificate verification.

Choose the installation’s network access method

Public HTTPS gateway

Expose the approved installation through your reverse proxy or API gateway with valid HTTPS, firewall and rate controls. Route the MCP endpoint, public discovery and JWKS routes, sign-in/consent pages, and OAuth endpoints. Keep authorization enabled. Configure the same canonical public origin for PUBLIC_APP_URL and BETTER_AUTH_URL. Issuer, audience and advertised endpoints come from that configuration, not untrusted Host headers.

Private connector or outbound tunnel

Use a connector approved for the chosen provider and customer network. OpenAI Secure MCP Tunnel is an OpenAI-specific option that forwards through OpenAI infrastructure; it is not a universal Claude/Copilot connector or an air-gap solution. Its operator needs appropriate OpenAI organization/workspace access, a tunnel runtime, outbound connectivity and a reachable installation. OAuth browser and token endpoints still need an approved route; tunneling the MCP endpoint alone does not establish the entire OAuth flow. This application does not create or start a tunnel automatically.

Follow OpenAI’s tunnel and OAuth network requirements. For Microsoft or other private connectors, verify the product’s transport and delegated-authentication support with the customer’s network administrator. A successful tunnel test for one provider proves nothing about another.

Local-only network

Keep the installation private and run the model, MCP client and optional bridge inside the authorized network. Login, token issuance, permission checks and tool execution stay with that installation. The optional CIMD fetch occurs only when a remote client metadata document is used; it can be disabled for disconnected environments.

Settings → AI connections → Network and connection diagnostics checks the configured URL, metadata, challenge and authorization endpoint from the installation server. It cannot prove reachability from an external AI service or a workstation. Test from the actual client network as well.

Administrator controls and data governance

Use the master control to enable access or disable and revoke the organization’s connections. In Organization access policy, choose all eligible members or selected user/group rules, limit capability scopes and tools, restrict readable procedures, and optionally require administrator approval for draft writing. Group grants combine and remain limited by the organization ceiling and the user’s current application roles. A procedure allow-list also disables new procedure creation.

Each authorized grant is bound to one person and one organization, even when many users share a published AI application identity. Consent selects access; registration alone grants no access to customer information. Connections expire after 30 days, access tokens after 15 minutes, with optional rotating refresh tokens bounded by the connection. Revocation or policy changes are checked on subsequent requests. Re-enabling a disabled organization does not restore revoked grants.

Connection management shows identity, status, capabilities, expiry and recent use. Administrators can approve or remove write access when required. Recent activity and the organization audit history show actions and outcomes. Audit records exclude tokens, client secrets, document contents and prompts. Failed OAuth requests include a request identifier for diagnostics.

Information returned to a cloud AI service leaves the installation for that service. Self-hosting alone does not make cloud-model processing local. Use approved clients, limit access, and review the client’s data handling and human-confirmation controls. Source documents and tool results are untrusted reference data, never authority to widen access. Structural validation does not prove safety or compliance.

Advanced manual OAuth compatibility

Clients without automatic identification can use Settings → AI connections → Advanced compatibility. Register exact HTTPS callbacks, or native loopback callbacks, and choose the client’s supported authentication method: public client with PKCE, client secret in HTTP Basic, or client secret in the token request body. Copy any one-time secret directly into the approved client. Private-key client authentication is supported through CIMD and standards-based client registration with public signing keys; it is not a shared API key.

For an enterprise agent shared by eligible users, an administrator selects Organization application. Registration itself grants no access: each user still signs in and consents within that organization, with their own current permissions. Administrators can revoke the managed application for the organization. Leave this option off for a personal compatibility registration.

Existing registrations remain valid. Old connection references and token resource bindings are migrated on startup. Changing the canonical public URL changes the OAuth authority and resource: plan a reconnect rather than redirecting old bearer tokens to a new host. Never enter tokens in callback URLs, logs, prompts or query strings.

Compatibility and current boundaries

The same six versioned tools discover/read procedures and formats, and create new procedure or standards drafts. A read-only authoring-format resource is also available. Tools are filtered by capabilities and organization policy. This release does not add existing-draft editing, publication, approval, tracked batch migration or automatic source-document extraction. Those are separate workflow phases, not authentication features.

Automated protocol tests and live provider-account acceptance are different evidence. Do not treat a successful metadata check as proof of a working provider connection. ChatGPT, Claude, Copilot Studio, Microsoft 365 tenant deployment, private tunnels and disconnected customer networks each need acceptance in the actual account and network. Provider-specific acceptance is not claimed until completed.

Connection errors: confirm the exact MCP URL and canonical origin, check organization access and membership, run diagnostics, then restart OAuth from the client. A client registered without a requested write scope needs a new consent request for that scope. If writing is pending, ask an administrator to approve that connection. For support, share a request ID and the error category, never credentials or tokens.

Open AI connections