Using the API from an AI
Two ways to give Fluenta to an AI: an agent with an API token and this documentation, or «Conectar con IA» for Claude and ChatGPT.
An AI can use Fluenta in two ways. Both act as you, with your role.
| Agent with an API token | «Conectar con IA» (Connect with AI) | |
|---|---|---|
| For | Any agent that makes HTTP calls | Claude and ChatGPT |
| Credential | An API token you create | A connection you approve with «Permitir» (Allow) |
| Organization's plan | Light and Pro | Pro |
| Coverage | The API's 62 operations | 38 tools |
| How it ends | «Revocar» (Revoke) | «Desconectar» (Disconnect) |
An agent that makes HTTP calls
- Create an API token in «Configuración» (Settings) → «Tokens de API» (API tokens). See Authentication.
- Keep it in the agent's secret settings, for example in an environment variable. Do not paste it into the conversation.
- Point the agent at the documentation as text:
| Address | What it holds |
|---|---|
https://docs.fluenta.io/llms.txt | The index: one line per page, to decide what to read |
https://docs.fluenta.io/llms-full.txt | Every page, whole, in one document |
https://docs.fluenta.io/openapi.json | The OpenAPI document: routes, parameters, permissions and answers |
https://docs.fluenta.io/en/docs/errors.md | One page as Markdown: add .md to its address |
The two llms files are in Spanish. Each page in Spanish is at
https://docs.fluenta.io/docs/errors.md, with the page's name in place of
errors.
Instructions to paste
Paste this into the agent's system instructions:
You have access to the Fluenta API over HTTP.
- Base: https://api.fluenta.io/v1/external
- On every request send the header
Authorization: Bearer $FLUENTA_TOKEN
The token is in the environment variable FLUENTA_TOKEN. Never write it in
an answer or in a log.
- Documentation: https://docs.fluenta.io/llms.txt is the index and
https://docs.fluenta.io/openapi.json has every operation.
- Start with GET /me, then GET /organizations. They answer the token's
permissions and the organizationId of each organization. Every other route
carries the organizationId in its path: there is no default organization.
- An id always comes from an earlier answer. Never invent one.
- Days are YYYY-MM-DD on the plant's clock, not UTC. The today field of
GET /organizations/{organizationId} is the plant's today.
- An error is { statusCode, code, message }. Branch on code and show message
to the person as it comes; it is in Spanish. Do not repeat a refused call
unchanged.
- On a 429 RATE_LIMITED wait the seconds of the Retry-After header before
you retry.
- Quote figures as they come. Do not recompute them.
- Delete nothing without asking for the preview and getting an explicit yes
from the person. A refused change changed nothing: do not say it went
through.«Conectar con IA»
Claude and ChatGPT connect over MCP, with no token. The address is:
https://app.fluenta.io/mcp- Open «Configuración» → «Conectar con IA». It shows the «Dirección del conector» (the connector's address) and the steps for each app.
- Paste the address into the app: in Claude, as a custom connector; in ChatGPT, as a custom MCP server with OAuth authentication.
- The app brings you to Fluenta. Review the request and press «Permitir».
Worth knowing:
- It comes with the Pro plan. A trial of Pro and an exempt organization have it too. The connection reaches only the organizations that have it.
- It is OAuth. You approve the connection in Fluenta with your dashboard session and choose its permissions: «Leer» (Read) always, «Gestionar el Programa» (Manage the Programa) if you tick it.
- It offers 38 tools. 23 only read: your identity and your organizations, the plant, the boards, the «Programa» (each module's production queue), the reference catalogue, the reports, the «Actividad de conexiones» (the record of changes) and the two delete previews. 15 change the «Programa» and the catalogue, and need «Gestionar el Programa»; two of those delete, and only with the confirmation of their preview.
- An API token does not work here. The MCP address accepts only the app's own connection: it rejects an API token like an invalid credential.
- The 23 read routes of 2026-11-12 have no MCP tool. An agent reaches them with an API token and this documentation. See the changelog.
«Desconectar», in Configuración, ends the connection at once.
Safety
- A token acts as you. It reaches every organization you belong to, with the role you have in each, read on every request.
- Prefer «Leer» only. Without «Gestionar el Programa», every write
answers 403
INSUFFICIENT_SCOPE. Tick it only if the agent must change the «Programa». - Give it a «Vencimiento» (expiry). Pick 30 days, 90 days or 1 year. A token set to «Nunca» (Never) works until you revoke it.
- One token per agent. You can hold up to 10 live tokens. With one per use, you revoke that one and leave the rest alone.
- Revoke it when you are done. «Revocar» ends it at once.
- An AI can follow instructions in the text it reads. Whatever the token can do, that text can ask it to do. One more reason to grant «Leer» only.
- Every change made through the API is in the «Actividad de conexiones», and
listActivityanswers it.