Skip to content
AI Primer
release

ReadMe launches Restless to repair failed agent API calls

Restless records agent API traffic and adds repair instructions to failed responses. Agents can replay calls through MCP, while missing information can be flagged as Linear issues or suggested documentation changes.

3 min read
ReadMe launches Restless to repair failed agent API calls
ReadMe launches Restless to repair failed agent API calls

TL;DR

  • ReadMe's Restless puts repair instructions and documentation links into failed API responses, according to the launch announcement.
  • Agents can search and replay their own calls through a hosted MCP server, as the setup walkthrough describes.
  • Agent feedback can become Linear issues or suggested documentation updates, according to the launch report.

The nicest implementation detail is in-process recovery lookup, with no network call in the error path. ReadMe says agents make up 72% of traffic to API docs. Its documentation updates extract reusable API knowledge from support conversations while leaving account details in the original thread.

Agent Recovery

Restless sits before the API's routes, leaving existing handlers in place.

Restless adds three pieces to 4xx and 5xx responses, according to ReadMe's SDK documentation:

  • A recovery object in JSON error bodies, containing the failed field, a hint, and a docs path such as /v1/invoices#currency.
  • An x-log-url header linking to the captured request.
  • An x-debug header containing a CLI command to inspect the call.

Teams can edit recovery guidance in the dashboard without changing API code. The next matching error receives the updated instructions.

MCP replay

Restless ties captured requests and responses to customer accounts, giving agents the history of their own calls.

The hosted MCP server also provides tools for changing product settings and lets users access the product from Claude or ChatGPT.

Feedback and doc approval

Restless adds a send_feedback tool to MCP so agents can report where integrations get stuck.

Agents are prompted to flag:

  • Incorrect descriptions.
  • Capabilities they cannot find.
  • Documentation that disagrees with API behavior.

Each note identifies the sending client and links to the originating conversation. The tool files a note without calling the API or changing product data.

Broken or missing features can be flagged as Linear issues, according to the launch follow-up. Documentation suggestions from resolved support conversations enter an inbox where a person can approve, edit, or reject them before publication.

SDK setup and redaction

Installation starts with npx restless init in the API repository.

The setup guide describes four steps:

  1. Locate the APIs and generate an OpenAPI specification in .restless/.
  2. Detect the runtime, install the SDK adapter, and wire the request middleware.
  3. Start local traffic capture and verify that requests reach Restless.
  4. Log in to claim the project.

Scanning runs through the developer's own Claude Code installation. Restless says source code is not sent to its servers during that step.

At runtime, the SDK redacts sensitive data on the API server before sending logs in background batches:

  • Headers: authorization, cookies, and API keys.
  • Request-body and query fields: passwords, tokens, secrets, Social Security numbers, and card numbers.

The CLI also checks the OpenAPI specification for custom authentication fields that need redaction.

Credits

Every project gets 1,000 free credits per month, covering log ingestion and the dashboard AI. ReadMe customers receive 4,000 additional monthly credits; further credits are sold in $100 increments.

Share on X