# Agent Door

> Plan how a SaaS or developer platform opens itself to AI agents, provider side - which agent classes may reach which
> surface, how each proves its identity and acts for a user, which execution path it gets, and how agent usage is
> counted. Answer a short interview, paste your robots.txt and any probe results; the browser sets the agent class by
> surface matrix, reads robots.txt per vendor token, runs the reachability gate, checks the agent-safe API basics and
> ranks the rungs. Metered runs write the identity and delegation plan, and the rollout. Built on the open-source skill
> @samber/agent-integration-surfaces, from samber/developer-platform-skills (MIT), by Samuel Berthe.

https://agent-door.skillsafe.ai/

## What is free
- The agent class by surface matrix: training crawlers, AI search, user-triggered fetchers, hosted browser agents, coding agents and fleets, across marketing, docs, signup, the app and the API - the skill's defaults plus your training calls (allow, deny or price).
- robots.txt read for 13 vendor tokens (GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot, Claude-User, Google-Extended, Google-Agent, PerplexityBot, Perplexity-User, CCBot, Applebot-Extended, meta-externalagent) with real group semantics and longest-match rules, held to the matrix: allow cells blocked, deny cells unstated, rules some fetchers ignore, content signals that contradict a training deny.
- The reachability gate from the CDN "block AI bots" preset, a signup CAPTCHA and pasted probe results; the agent-safe API checklist with an owner for each failure.
- The rungs A to G - measurement tags, docs markdown and Context7, agent credentials and delegation, skill plus --json CLI, MCP server, provision-then-claim signup, discovery files - deleted and re-ranked by your answers, with the execution path per class. Plus a robots.txt that states the matrix, the matrix (.csv), the checks (.csv) and everything (.json).

## The metered lanes
Two tasks on gpt-terra, metered per run, routed by `task`. Input string fields: `task` (plan or rollout), `product`, `agents`, `carry` (rollout only: the plan decisions) and `facts`.
- `plan` replies with `summary`, `identity` (per class: identified_by, authority, credential, scope, lifetime, sponsor, revocation, approval), `paths`, `signup`, `rungs` and `caveats`.
- `rollout` replies with `summary`, `gate_fixes`, `measurement` (tags, baseline, funnel), `steps`, `exit` and `caveats`.
The page holds every reply to its own facts: an identity row per class with an expiring credential, a sponsor and per-agent revocation, never a human's key, crawlers and browser agents verified by IP list or signature, money actions approved outside the agent's channel, the page's paths, signup decision and rungs; a fix per failed gate, the agent-class field with unknown as a value, a baseline, a funnel, agent share as a floor, one step per scheduled rung and at least three exit criteria.

## Links
- API tutorial: https://agent-door.skillsafe.ai/api.html
- Source skill: https://github.com/samber/developer-platform-skills/tree/main/skills/agent-integration-surfaces