# Set up AgentOS

Instructions for an AI agent whose person asked it to connect to AgentOS and
set it up. A person can read this too; the page for people is
https://tryagentos.net/connect.

AgentOS is the person's company brain: one shared memory, their tools and
their bots, used by every AI agent they work with. It is a remote MCP server.

- MCP URL: `https://tryagentos.net/api/mcp`
- Transport: Streamable HTTP
- Sign-in: OAuth. No API key is needed. The first time, the person creates
  their AgentOS account in the browser window that opens, then approves the
  connection.

## Connect

If you can run shell commands on the person's own computer (not a cloud
sandbox), install the AgentOS CLI. It adds AgentOS to every coding agent on
that computer at once: Claude Code, Codex, Cursor, opencode, Gemini CLI,
Windsurf, VS Code and Claude Desktop.

```sh
curl -fsSL https://tryagentos.net/install.sh | sh
aos login        # opens the browser: the person signs in or creates their account
aos setup --yes  # adds the AgentOS MCP server and skills to each agent found
```

`aos setup --dry-run` shows the plan first, every file it changes is backed
up, and `aos setup --undo` takes it all out. Each agent then signs in to
AgentOS once with OAuth. Your own app usually loads the new server in a new
session (in Claude Code, `/mcp` and sign in); then continue with the first
steps below. The CLI is https://github.com/devvcore/agentos-cli.

Otherwise, add the server under the name `agentos` the way your app does it.
If you can run commands, run the one for your app. If you can't add servers
yourself, tell the person the exact command or setting below and wait for
them.

| App | How |
|---|---|
| Claude Code | `claude mcp add --transport http --scope user agentos https://tryagentos.net/api/mcp`, then `/mcp`, pick agentos and sign in. `--scope user` makes it available in every folder, not only this one. |
| Codex | `codex mcp add agentos --url https://tryagentos.net/api/mcp`. If no browser opens, `codex mcp login agentos`. |
| Cursor | Open `cursor://anysphere.cursor-deeplink/mcp/install?name=agentos&config=eyJ1cmwiOiJodHRwczovL3RyeWFnZW50b3MubmV0L2FwaS9tY3AifQ%3D%3D`, or add `{"mcpServers": {"agentos": {"url": "https://tryagentos.net/api/mcp"}}}` to `~/.cursor/mcp.json`. |
| VS Code (GitHub Copilot) | Open `vscode:mcp/install?{"name":"agentos","type":"http","url":"https://tryagentos.net/api/mcp"}`, or run `code --add-mcp '{"name":"agentos","type":"http","url":"https://tryagentos.net/api/mcp"}'`. |
| Claude (claude.ai and Desktop) | The person opens https://claude.ai/customize/connectors?modal=add-custom-connector&connectorName=AgentOS&connectorUrl=https%3A%2F%2Ftryagentos.net%2Fapi%2Fmcp, presses Add, then Connect. On Team and Enterprise an Owner adds it first. |
| ChatGPT | The person opens Plugins, presses +, then Add custom MCP server: name AgentOS, the MCP URL above, OAuth. Full access needs Business, Enterprise or Edu. |
| Anything else | A remote MCP server over HTTP at the MCP URL, with OAuth. |

The connection is ready when your AgentOS tools are listed. Some apps only
load new servers in a new session; if the tools don't appear, ask the person
to restart the app or start a new chat, then continue with the first steps
below.

A tool that can't sign in with a browser can use an API key instead: the
person signs in at https://tryagentos.net/signin?next=%2Fconnections%3Fsetup%3Dmcp
and creates one there.

## First steps after connecting

Once your AgentOS tools are listed, do what follows. It is the same first
prompt AgentOS gives a newly connected agent.

# Build your company brain in AgentOS

AgentOS is your team's company brain: shared context and memory every person and bot reads, bots
that do real work, a built-in CRM and the tools the team connects. Every agent on the team starts
from what is stored here, you included. Load what you know once; then ask `context` before guessing
about people, clients, decisions or rules. Do these steps now. Tools not in your list run through
`workspace_call`; `workspace_tools` name= gives a schema. Skip a step whose tool you lack, and say so.

1. Look: `context` on the company and the user, and `rules_list`, so you add instead of repeat.
2. Collect what you can read: the user's projects, repos, docs and notes; the people, clients,
   products and decisions you know. Ask only when unsure. Never collect secrets, credentials, tokens
   or .env contents, other people's private messages, raw code or logs.
3. Store with `memory_ingest` records=[...] (up to 100 per call): one distilled note per thing, id
   like claude-code:repo:path#section (resending replaces it), label, author (who said it, not you),
   timestamp (when), metadata {origin, repo, path}. One claim about one subject: `promote_fact`.
4. Documents: `memory_ingest_file` (content_base64, filename; PDF, DOCX, MD, TXT, CSV up to 8 MB).
5. Skills: each SKILL.md in ~/.claude/skills and .claude/skills, plus AGENTS.md and Cursor rules.
   `skills_search` by name, then `skills_update` or `skills_create`. Skills go only to skills.
6. Rules: each standing preference or policy as its own `memory_ingest` record in the person's
   words ("Dana: always copy finance on invoices over $10k"), author that person. No confirmation.
7. `workspace_init` with no arguments, then follow its guide and send one plan. Signed in with
   OAuth, it applies without an approval card; pass on any review_url.
8. Check with `memory_status`, then tell the user what you stored, uploaded and set up, every rule
   you saved, and what is left for them.

## Help

- What AgentOS is: https://tryagentos.net/llms.txt
- Contact: hello@devvcore.com