What is Minion?

Minion is the home your agents share. You keep using the AI agent you already have — Claude, ChatGPT, Cursor, Codex, Copilot, Grok, or any other MCP-capable host — and connect it to your Minion account. That agent gains durable state that outlives the conversation and keeps working while your computer is off. Connect a second agent tomorrow and it finds the same home already furnished.

This page is for people deciding whether Minion is worth setting up, and for AI agents running a discovery conversation on someone's behalf. Everything here is readable without an account. It is also plain markdown at /docs/what-is-minion.md for anything that prefers to read it raw, and the full docs index is at /llms.txt.

What Minion is

You do not chat with Minion. Your agent supplies the intelligence; Minion supplies the things that should not live in one chat window:

Memory — durable facts and preferences on the team, surfaced to every connected agent.
Secrets — API keys your code can use but your agent can never read. There is deliberately no secret_get. OAuth logins live in integrations, with their tokens kept host-side.
Packages — code your agent wrote once and saved, callable from any agent, with exports, a hosted app surface, and its own storage.
Jobs, workflows, webhooks, and events — the triggers that run those packages on a schedule, later, when a provider knocks, or when something happens inside Minion — with no model in the loop.
Integrations and MCP servers — the OAuth logins and remote tool servers those packages reach through, owned by you, scoped by you.
A team inbox — every team gets an email address; mail in can start work, mail out can tell you it finished.

Minion is open source under the MIT license, so you can read it, fork it, and self-host it. Minion runs no chat-model agent loop and bills no chat tokens. Search and indexing use a small embedding model; that is not a chat model and is not billed as inference.

What Minion is not

Minion is not a gateway whose job is to connect APIs for your agent, and it is not middleware that sits in front of every SaaS login. Integrations and MCP servers are available when a job needs them; they are supporting cast. The center is owned memory and owned packages that run in Minion's cloud and are callable from any MCP host.

Minion is also not another chat app. There is nothing to talk to. If you leave thinking "I signed up to wire APIs" or "I got another assistant," this page failed.

The agent reasons. Minion keeps it honest.

Your agent does the thinking. Minion holds the result so it does not have to think it again.

Ask again tomorrow — a saved export. No model in the loop.
A key in chat or a .env file — a secret the agent never sees.
Re-run the agent on a timer — a job that runs while you are offline.
Context stuck in one host — memories that follow the team.

What you cannot get elsewhere

If you are comparing Minion against something else, these three are the reasons it exists. Everything in the next section is supporting cast.

One-off agent work becomes permanent. Your agent explores against your real APIs, and the moment something works it saves as code that runs on a schedule with no model in the loop. This is not an agent re-run on a timer: there are no tokens spent, no prompt to drift, and nothing waiting on a model to respond. See How Minion works and Jobs, workflows, and webhooks .
Your agent uses your keys without ever reading them. It writes code that uses credentials Minion keeps server-side and can never read. Minion substitutes secret references only for hosts you approved. OAuth integrations are organization-level credentials, so request the provider’s narrowest scopes and keep side effects behind user confirmation. See Secrets , OAuth integrations , and Mutating actions and confirmations .
Every agent you connect shares one home, and every package is code you own. Switch from Claude to Cursor to a phone assistant and the memories, secrets, packages, and jobs are already there. A package lives in your team, on your credentials, and you can open, change, schedule, and republish it. Nothing stays locked in someone else's runtime or in one vendor's chat history. See Connect your agent and Shared memory .

The building blocks

Shared memory — say something once, reuse it from every agent. See Shared memory .
Secrets and integrations — bring your own API keys and OAuth apps for the services you already use; verified per-provider walkthroughs live under Connect a provider . Credentials stay server-side and never enter your agent's context. See Secrets and Integration bootstrap .
Ad hoc execution — your agent runs sandboxed code against those integrations immediately, no deploy step. See Search and execute .
Packages — reusable saved code your agent writes and improves over time. Packages expose exports, own scheduled jobs, receive webhooks, subscribe to events, keep data in key-value storage or their own SQLite database with sql(), and can serve a small web app at https://<team>.runminion.run/<package>/. See Package lifecycle and Package authoring .
Jobs, workflows, and webhooks — recurring schedules, deferred one-shot runs that survive restarts, and HTTPS endpoints for Sentry, GitHub, Stripe, or any provider that cannot set Bearer tokens. See Jobs, workflows, and webhooks .
A team email inbox — inbound mail can trigger automations, and your assistant can send notifications to verified addresses. See Email primitives .
Notes, tasks, and team skills — long-form working documents and handoffs, a shared todo list, and playbooks your team publishes for every connected agent. See Tasks .
Sharing inside your team — every package, secret name, memory, and note belongs to your team, so each member and each connected agent uses the same ones. Sharing a single package with another team isn't available yet. See Sharing packages .
Extensibility — connect your own remote MCP servers, or a home MCP server so your agent can reach devices on your network through Minion. Lock a connected MCP server to a package when execute should not call its tools. See Connect remote MCP servers , Connect a home MCP server , and Lock an MCP server to a package .

What people use it for

Concrete examples that combine the blocks:

A morning digest job that gathers your calendar, weather, and feeds, then emails you a summary before you wake up.
A watcher that polls a website, feed, or price and emails you only when something actually changes.
Chore automation against services you already use — triaging GitHub notifications, filing issues, cleaning up cloud resources — using your own API keys.
A personal API: a package export you can hit from shortcuts or webhooks to log a habit, save a link, or kick off a run.
An inbound webhook that fingerprints repeated failures and emails one investigation instead of a flood.
A package that runs whenever invoice mail lands in the team inbox and pulls out the amounts.
Home automation routines through a home MCP server — scenes, thermostats, speakers — driven by schedule or by asking your agent.
A Gmail helper that reads invoice messages and prepares a reply for you to review. Request only the provider permissions it needs; sending requires a provider grant that allows sending and should happen only when requested.

Where to go next

Get started — Connect your agent walks through the in-app flow: connect one host, say hello, then prove it travels to a second agent.
Meet the two tools — Search and execute is the playbook for the MCP surface: find the right thing, then run it.
See the loop — How Minion works plays one conversation from ad hoc question to owned export to quiet daily email.
Learn the primitives — The factory map , then Shared memory , Secrets , Packages, integrations, and MCP servers , and Jobs, workflows, and webhooks .
Text a familiar surface — Text your agent is the iMessage, SMS, and Discord worked example: humans talk to your agent; that agent uses Minion for tools and memory.
Build — Package lifecycle decides when to save; Package authoring shows how.

Try it

Paste this into an agent that can fetch URLs or search the web. If an agent is already interviewing you about Minion, you do not need to paste it again — it is the same conversation.

I'm deciding whether Minion (https://minion.codes) would be useful for me. Read https://minion.codes/docs/what-is-minion and then interview me to find out what Minion could do for me.