← All docs

If you're not sure, Sanction it

Ask for one human approval from your AI tool. Review the exact proposed action, then approve or deny it. You do not need to write a policy first. Existing hard blocks still apply.

Sanction is free for individuals. Start with the harmless test below: it requests permission and executes nothing. Connecting Sanction does not automatically intercept the host's other tools.

Per-surface install paths (ChatGPT, Codex, Claude, Cursor, Slack), including disconnect: Install Sanction. ChatGPT and Codex also have a setup and disconnect guide.

Choose your AI host
Full connection reference, troubleshooting, and disconnect

If you're not sure, Sanction it

Ask for one human approval from your AI tool. Review the exact proposed action, then approve or deny it. You do not need to write a policy first. Existing hard blocks still apply.

Sanction is free for individuals. Start with the harmless test below: it requests permission and executes nothing. Connecting Sanction does not automatically intercept the host's other tools.

Per-surface install paths (ChatGPT, Codex, Claude, Cursor, Slack), including disconnect: Install Sanction. ChatGPT and Codex also have a setup and disconnect guide.

1. Choose your host

Create your Sanction account first. OAuth connections require a verified account and a wallet you own or administer. Select an active agent or create a dedicated test agent on Sanction's consent screen, then connect it. Never paste an agent key into chat.

Use this Streamable HTTP URL for approval-focused OAuth connections:

https://getsanction.com/mcp/approvals

This connection exposes eight existing tools and excludes execution-token issuance and vault credential retrieval. It is a custom connection, not evidence of a Sanction directory listing.

Which MCP profile?

Sanction serves two MCP profiles from one decision engine. Both stay available; pick by what the agent needs.

ProfileEndpointAuthTools
Approvals, for AI hosts/mcp/approvalsOAuth; API key also works8
Full wallet, for developers/mcp, or npx sanction-mcp (stdio)Agent API key10
  • Approvals profile (https://getsanction.com/mcp/approvals): authorization requests (sanction_authorize, sanction_authorize_provision, sanction_authorize_tool, sanction_authorize_capability), decision checks (sanction_check_authorization), sanction_wallet_status, and usage/business-outcome logging (sanction_log_tokens, sanction_log_outcome). No execution-token issuance or vault credential retrieval.
  • Full wallet profile (https://getsanction.com/mcp): the same eight tools plus sanction_request_execution (scoped execution tokens) and sanction_inject_credential (credential retrieval from the encrypted vault).

These URLs are MCP protocol endpoints for hosts, not pages to open in a browser. Both profiles are cooperative: the host must consult Sanction and honor its answer. Enforcement of upstream tool calls applies only to traffic routed through the MCP broker. The /.well-known/mcp.json manifest lists both profiles.

Host setup and evidence

HostSetupSanction evidence as of October 22, 2026
ClaudeCustomize → Connectors → Add custom connector. Enter the URL above, choose automatic registration when offered, and complete Sanction consent.Production synthetic request, approval, exact redemption and consumed-grant refusal observed October 1. No external action executed.
CursorAdd the OAuth configuration below to your MCP settings and authenticate.Configuration follows Cursor documentation; Sanction lifecycle test pending.
CodexAdd the HTTP server using the commands below, then sign in.Configuration follows Codex documentation; Sanction lifecycle test pending.
Grok (consumer)Connectors → New Connector → Custom. Enter the URL above and complete the offered authentication.Custom /mcp/approvals OAuth completed October 6: consent identified Grok, eight tools enabled, one sanction_wallet_status read. Approval lifecycle still pending. Not a directory listing.
Grok BotIn chat, ask the Bot to add https://getsanction.com/mcp/approvals. Complete Sanction OAuth. On desktop you can also use Connect apps → Add a custom connector → Remote HTTPS.Unauthenticated /mcp/approvals challenges with WWW-Authenticate resource metadata (production, October 22). A Bot session reached OAuth sign-in with no custom headers. Consent completion, tool list, and approval lifecycle are unverified. Not a Connect-apps listing.
ChatGPTNo verified public listing. Use the ChatGPT/Codex setup guide when a listing is published. A local Codex connection does not install ChatGPT web.No verified public listing or completed ChatGPT lifecycle test. Package candidate only.

Cursor

Merge this entry into your project .cursor/mcp.json or global ~/.cursor/mcp.json; preserve existing servers:

{
  "mcpServers": {
    "sanction": {
      "url": "https://getsanction.com/mcp/approvals"
    }
  }
}

Use Cursor's authentication control when prompted. This config contains no key. The Cursor Directory community listing links to the standalone plugin, which uses the full MCP profile and requires an agent key. This is a separate installation path from the OAuth setup above.

Codex

codex mcp add sanction --url https://getsanction.com/mcp/approvals
codex mcp login sanction

Use Codex 0.147 or later; older versions can fail at sign-in with a missing-issuer error (details). Use an unused server name if sanction is already configured. Approve the intended agent in Sanction, then restart or refresh the host's MCP connection as needed. Local Codex clients share MCP configuration; ChatGPT web does not read it.

Grok Bot, xAI API, and Slack

Grok consumer, Grok Bot, and the xAI API are separate connection paths. The Grok Bot guide describes the in-chat custom URL, OAuth-only sign-in without a key, and shared-account boundaries. Consumer custom-MCP support does not prove a Grok Bot listing or a completed Bot approval lifecycle.

Slack is where a person can receive and decide a request. Configure Sanction for Slack separately; adding MCP to an AI host does not install the Slack app. Use the email approval link if Slack delivery is not configured.

Official host references: Claude, Cursor, Codex, Grok, ChatGPT plugin submission. Host menus and account availability can change; these instructions were checked October 22, 2026. Grok Bot desktop setup was updated November 12 using the official guide.

2. Check the connection

Ask your host:

Call only sanction_wallet_status with empty arguments. Summarize the budget and pending-approval count. Do not create an authorization or display credentials.

A successful read proves the connection, not enforcement of other tools. If the host cannot find Sanction or asks for authentication, fix that before proceeding.

3. Request your first approval

Copy this prompt into the connected host:

Ask Sanction for my approval to run sanction.demo.approval with arguments {"synthetic":true,"execute":false} and no server. Set require_approval to true and approval_reason to "Tests the approval flow only; executes nothing." Make one sanction_authorize_tool call, show the request ID and any approval link, then stop. This is a synthetic test: do not execute any action, retry, or poll automatically.

Expected: authorized: false, status: escalated, and next_action: wait. A hard policy denial is also a valid result; it is not overridden by asking for approval. Open the approval notification or your approval inbox, review the exact arguments, and approve or reject that test request.

If Claude's tool schema lacks require_approval, open the existing connector's options and choose Refresh tools list, then retry discovery. This was needed after the schema update in the October 1 smoke. Do not silently omit the flag.

4. Check once, then redeem exactly once

After reviewing the request in Sanction, paste this into the same AI conversation that created it:

Check the synthetic sanction.demo.approval request from this conversation once using sanction_check_authorization and its original request ID. If the original request ID or exact submitted fields are missing or ambiguous, ask me to identify the request and stop; do not guess or create a new request. If the result is approved, has next_action: retry_with_grant, and includes an active grant, make exactly one sanction_authorize_tool call with the original tool, server (including its omission), arguments, require_approval, and approval_reason unchanged, plus that grant_id. Otherwise report the status and stop, including if pending, denied, expired, or unclear. After redemption, report the status, authorized, next_action, and request ID, without displaying the grant. Do not execute the synthetic tool or any external action, poll, retry again, or request fresh approval—even if redemption returns authorized: true and next_action: proceed.

After you decide, ask the host to call sanction_check_authorization once with the request ID. A denial means stop. Approval alone is not execution permission: next_action: retry_with_grant means repeat the original authorization with the same tool, server and arguments, plus the returned grant_id.

Only a redemption returning authorized: true and next_action: proceed permits that attempt. For this synthetic test, report the result and execute nothing even after redemption. Never automatically request fresh approval after a refusal, an ambiguous result, or a consumed grant. A yes in chat is not a substitute for checking, and it cannot override organizational policy or a hard denial.

A decision record proves the authorization decision, not that an external action ran. sanction_log_outcome records business outcomes, not execution receipts. If you cannot tell whether an approved action ran, treat it as unknown: do not retry automatically, and do not treat expiry as proof that nothing happened.

Tool grants bind the reviewed arguments and expire. Redemption checks the grant; it does not re-evaluate the current tool-policy ladder. OAuth and decision contract details explain the boundary and remaining test gaps.

Disconnect or get help

For OAuth, revoke the host at MCP connections, then remove the connector in the host. Revocation stops future requests through that connection. Earlier decisions and admitted work keep their recorded status.

For the Cursor community plugin, open it under Agents and select Rotate key, then remove the plugin from Cursor. Every client using the old key loses access when you rotate it. Removing the plugin alone leaves the key valid.

Report setup problems through Sanction support. Include the host and failed step, but no keys, tokens, grants, or private action arguments. For enforcement, route supported traffic through the MCP broker; a cooperative connection cannot govern calls that bypass it.