GETTING STARTED
Integrate with an AI agent
These docs are written for people. This page is written for the agent doing the work.
Most Agentifys integrations are now written by a coding agent rather than typed by hand. An agent reading this site gets React markup, navigation and a theming playground wrapped around the two facts it needed. So the same material is published in formats an agent can consume directly: a spec, a single-file reference, and a skill. And for the setup half of the job, an agent does not have to read anything at all. It can call the platform directly over MCP.
The four files, and one endpoint
| Field | Type | Description |
|---|---|---|
/llms.txt | map | A short index: what exists and where it lives. The file to fetch first. |
/llms-full.txt | reference | The entire integration surface in one self-contained file: endpoints, signing, SSE events, widget, tools, errors. An agent can integrate end to end without fetching anything else. |
/openapi.json | OpenAPI 3.1 | The public /v1 API, 19 operations. Generate a client from it. Models the per-route identity transport structurally, which is the thing integrators most often get wrong. |
/agentifys.skill.md | Agent Skill | Task-shaped instructions: a decision tree, the integration checklist in order, how to verify each step actually worked, and a troubleshooting table keyed on the symptom you actually see. |
/mcp | MCP server | Not a file. A live endpoint. Connect an agent with a scoped eat_ admin token and it gets twelve tools that create and configure projects: prompt, user schema, tools, policies, MCP servers, limits, scheduling. See the MCP provisioning guide. |
/admin routes: the console API is not public./v1 API use the publishable era_ project key. The MCP server uses an eat_ admin token, which is the opposite kind of credential: secret, server-side only, never in browser code and never in a committed file. Both ride Authorization: Bearer, which is exactly why agents mix them up, the skill file's first section exists to stop that.Give them to your agent
# Claude Code. Drop it in your project mkdir -p .claude/skills/agentifys curl -sSL https://agentifys.ai/agentifys.skill.md \ -o .claude/skills/agentifys/SKILL.md # Or hand the whole reference to any agent in one shot curl -sSL https://agentifys.ai/llms-full.txt
The skill file is a single markdown file with YAML frontmatter, so it works anywhere skills are supported, and reads fine as a plain document anywhere else. Nothing about it is Claude-specific beyond the format.
The files tell an agent how to integrate a project. If you also want it to create one, connect it to the MCP server as well, then the same agent writes the prompt, declares the user schema, adds the tools and the policies, and reports what a human still has to do by hand:
claude mcp add --transport http agentifys https://agentifys.ai/mcp \ --header "Authorization: Bearer eat_..." \ --scope user
A prompt that works
Agents integrate better when they are told which of the two shapes you are building before they start, because that single choice decides the rest: identity mode, where the key lives, and whether your allowed-origins list must stay empty.
Integrate Agentifys into this product. Read https://agentifys.ai/llms-full.txt first. It is the complete integration surface in one file. The OpenAPI spec is at https://agentifys.ai/openapi.json. Our situation: - we serve signed-in users, so use signed identity - our backend will proxy the calls (the browser never holds the key) - we want the widget, one read tool, and a weekly scheduled report
If you are a person
Start at the integration walkthrough, which carries one realistic scenario from first request to a working scheduled task, and links the reference pages at each step. The quickstart is faster if you only want the widget on a page.