MCP and OAuth
The Streamable HTTP endpoint is https://ink.ego.ist/mcp. ChatGPT, Claude and Claude Code setup use this same endpoint.
Authorize a client
Discover the protected resource metadata and authorization server metadata. Public clients use authorization_code with S256 PKCE, the MCP resource URL as the resource audience, and either dynamic client registration (DCR) or a client ID metadata document. Discover the authorization, token, registration and revocation URLs from this metadata.
The person signs in and explicitly consents. The only scope is ink.rewrite, covering Ink account status, available models, writing tools and owned jobs and results. It grants no Passport memory access. Signup and subscription handoffs do not authorize a client. Refresh tokens rotate; use the newly returned token after each refresh.
Registration allows 30 requests per IP per hour. MCP allows 120 requests per minute per connected user and client. New rewrite jobs share 10 per minute per user across clients. Wait for Retry-After when supplied.
Tools
ink_get_status, ink_rewrite_text, ink_write_text, ink_get_job, ink_get_result, ink_cancel_job
Call ink_get_status to check setup and available words. Use ink_rewrite_text with text, a UUID idempotency_key and optional settings. Strength defaults to balanced; light and strong are also available. Context is general, work, casual or academic; formality is auto, more formal or less formal.
Rewrite input allows 20,000 words and 200,000 characters. Optional locks and unlocks contain verbatim input strings. Each array allows 100 strings with 20,000 total characters across both arrays. Locks preserve wording; unlocks release automatic highlights, including nested spans. Missing or invalid entries return a validation error. Changing text, settings, mode, profile or locks requires a new idempotency key.
Tools return setup_required or quota_exceeded with a link to plans. Relay the message and link to the person. Policies describes terms_required and feedback sharing. Feedback tools are available only to the Ink widget on ChatGPT and Claude, not as model-visible tools.
Meaning checks
Ink silently checks rewrite meaning with OpenAI when configured. Every changed candidate is checked, including retries. A meaning finding can use the section's single shared retry, with source-quote hints. Deterministic guards take precedence when choosing a candidate, then fewer material or critical findings win, with the retry winning a tie. Light and balanced deliver persistent meaning problems silently. Strong keeps and refunds a source section through sections_kept only when the chosen candidate still has a critical added claim, changed fact or speaker problem. Material findings are delivered silently; dropped findings can trigger a retry and affect selection but do not keep source by themselves. Unavailable checks fail open. Older saved policies keep their behavior.
Write from a prompt
Use ink_write_text with prompt and a UUID idempotency_key. Put all instructions in the prompt. Prompts allow 6,000 words, with up to 4,000 output words. Ink reserves available words up to the output cap, charges actual output words and refunds the remainder. Failed or cancelled prompt jobs refund the full reservation. Use a fresh key for another version. Rewrite fidelity checks apply only to rewrites.
Follow progress
Poll ink_get_job and ink_get_result. Wait two seconds while queued or starting_model, then one second while writing, rewriting or validating. Cold starts can take a few minutes. blocks contains ordered completed previews; missing indices are pending. progress has phase, blocks_done and blocks_total. Partial blocks are not final output. Warnings and change_summary are final only.
completed, completed_with_warnings and needs_review are finished results. Deliver warnings with the output. For details_added, first_person_added or sections_kept, do not retry automatically. failed has no usable result; cancelled is stopped. Use ink_cancel_job to request cancellation, then poll until terminal. After text expiry, output is absent and blocks is empty.
On details_added, deliver the output with its warning listing added details. The message is "Check added details before use: ...". On first_person_added, deliver the output with its warning listing added first-person words. The message is "Added first-person wording: ..." or, when the list is absent, "Added first-person wording. Check the rewrite.". Light and balanced deliver persistent first-person additions with completed_with_warnings; strong keeps the affected source section and refunds it through sections_kept. Read the first-person rules and quotation limit.
Revoke access
Revoke through the discovered revocation endpoint, or choose Disconnect in Connections. Disconnect removes the assistant's Ink access. It does not cancel a subscription.