Connect with MCP

RigorDesk exposes its review workflows through the Model Context Protocol (MCP): start multi-model reviews, poll them, and read verdicts without leaving Claude, ChatGPT, or Codex. Connect once with OAuth — no API keys to paste — and the assistant calls RigorDesk tools mid-conversation.

Server URL

https://api.rigordesk.com/mcp

Connect from Claude

Claude supports any remote MCP server as a custom connector, on every plan including Free:

Open Customize → Connectors.
Click +, then Add custom connector.
Paste the server URL above and authenticate in the browser with your RigorDesk account.

Connect from ChatGPT

On a Business, Enterprise, or Edu workspace with developer mode:

Open workspace settings and enable developer mode.
Add a custom MCP app with the server URL above.
Sign in with your RigorDesk account when prompted.

Free and Plus users can only use listed apps, not custom server URLs.

Connect from Codex

In the Codex CLI or desktop app, add RigorDesk as a remote MCP server:

[mcp_servers.rigordesk]
url = "https://api.rigordesk.com/mcp"

Then run codex mcp login rigordesk to complete the browser OAuth flow. For scripted, non-interactive use, set a bearer token instead:

[mcp_servers.rigordesk]
url = "https://api.rigordesk.com/mcp"
bearer_token_env_var = "RIGORDESK_API_KEY"

with RIGORDESK_API_KEY set to a Pro API key (dbt_…) from Settings → API Keys. Never commit the key or paste it into shared configs.

Available tools

Tool	Use
`get_usage`	Tier, period spend versus budget, and prepaid credit balance
`list_models`	Model catalog: IDs, tiers, and context lengths for panel building
`list_assemblies`	Assemblies you own or that are public
`get_assembly`	Inspect one assembly's panel configuration
`create_assembly`	Create a panel from model IDs (free, same validation as the app)
`update_assembly`	Fix an owned panel; only supplied fields change
`start_review`	Start an async, metered review (up to 50,000 characters)
`get_review_status`	Lightweight progress check without the transcript
`get_review`	Verdict and summary, optionally with the full transcript
`list_reviews`	Your previous reviews, filterable by status and assembly
`cancel_review`	Stop a running review

A typical flow:

get_usage (optional, when spend matters)
list_models (to pick panel model IDs)
list_assemblies (or create_assembly when nothing fits)
start_review(content, client_request_id, [assembly_id])
get_review_status(review_id)
get_review(review_id, include_transcript: true)

start_review needs a client-generated client_request_id. Reuse it when retrying the same intended review — RigorDesk returns the original instead of starting a duplicate. Starting a review spends your normal allowance or prepaid credits; creating panels and reading results is free.

Large transcripts are budgeted: when a response reports has_more with next_round/next_turn_offset, continue with include_transcript: true plus those values until everything is retrieved.

Billing

MCP usage bills exactly like the web app and the REST API: one metered run per start_review, deducted from your subscription allowance first, then prepaid credits. See pricing & billing for tiers and pricing to subscribe or top up.

Troubleshooting

OAuth discovery fails — the client must support Streamable HTTP transport and OAuth 2.1 with PKCE. Re-add the connector with the exact server URL above.
401 responses — sign in again; tokens expire. Confirm you approved the consent screen for the right RigorDesk account.
Review won't start — call get_usage to check your allowance, then list_assemblies to confirm the panel is visible to you. Errors carry stable codes (BUDGET_EXHAUSTED, ENTITLEMENT_DENIED, CONCURRENCY_LIMIT).
Review still running — poll get_review_status every 5–10 seconds. Don't re-call start_review; use cancel_review to stop it.