---
name: cfoai-mcp-setup
description: Set up cfo.ai for this agent, from account access through MCP authentication and a verified first read. Use when the user asks to connect to cfo.ai or onboard to Ari.
metadata:
  author: cfo.ai
---

# Set up cfo.ai

Connect this agent to Ari, the cfo.ai CFO, in the user's intended workspace.
Canonical guide: https://cfo.ai/skill.md. This is a Markdown instruction file,
not a shell command. Read it and carry out the applicable setup steps.

## 1. Establish account and workspace access

- Existing users sign in at https://app.cfo.ai/login.
- New users start at https://cfo.ai/start. This opens signup or the access-request
  form depending on availability. If access is pending, explain that approval
  is needed before MCP authentication can finish; resume once access is granted.
- Help the user reach account creation and any required plan setup in the app.
  Let them complete sign-in, verification, payment, and authorization in the
  browser. Do not ask them to paste passwords, verification codes, payment
  details, or tokens into chat.
- Use the workspace the user names. Do not create another account or workspace
  if they already have one. If the intended workspace is missing, direct them
  to its owner for access.

## 2. Connect the MCP server to this agent

Identify the current client and inspect its existing MCP configuration first.
Reuse a matching server instead of adding duplicates; preserve other servers.

- Name: `cfoai`
- URL: `https://api.cfo.ai/mcp`
- Transport: remote Streamable HTTP
- Authentication: OAuth through the client; no manually supplied API key

### Claude Code

```bash
claude mcp add --transport http cfoai https://api.cfo.ai/mcp
```

Restart Claude Code if needed. Open `/mcp`, select `cfoai`, and choose
**Authenticate**. Let the user complete authorization in the browser.

### Codex CLI

```bash
codex mcp add cfoai --url https://api.cfo.ai/mcp
codex mcp login cfoai
```

Let the user complete the browser authorization. Reload the session if the
new server's tools are not yet visible.

### Other agents and desktop clients

Use the client's supported MCP configuration command, settings, or config file
for a remote HTTP server named `cfoai` at the URL above, then use its OAuth
sign-in flow. Check that client's documentation for its exact configuration
format; do not assume every client accepts the same JSON or CLI commands.

If this agent cannot configure its own client, give the user the exact setup
steps for their client and resume when the server is available. If the client
cannot support remote HTTP MCP with OAuth, explain that limitation and suggest
a compatible client. Do not substitute an invented key or token workaround.

## 3. Select the workspace and prove the connection

1. Confirm that the server is connected and its tools are visible. Discover the
   available tool names, schemas, and initialization instructions from the
   server rather than assuming a fixed tool count or argument structure.
2. Call `list_orgs` to identify accessible workspaces. A single workspace may
   resolve automatically. If several are available and the user's intended
   workspace is unclear, ask them to choose. Use its ID or slug as `org_id` on
   subsequent calls according to the tool schemas.
3. Run a read-only inspection in that workspace. The product docs describe
   tools such as `inspect_reference`, `inspect_variables`, `inspect_dimensions`,
   and `inspect_table_blocks`; choose an available one and follow its schema.
   An empty workspace is a valid result, not a reason to create test data.
4. Report the client/server name, authenticated connection status, workspace,
   and the read-only call that succeeded. Configuration alone is not proof.
   Do not say setup is complete while authentication or verification is pending.

If authentication fails, retry through the client's OAuth flow and check that
the user signed into an account with workspace access. If tools are missing,
reload the client and check the server configuration. If still blocked, report
the failing step and the next action without exposing credentials.

## 4. Continue with the user's task

Once verified, follow the server's initialization instructions and read
https://docs.cfo.ai/agent-instructions.md for product operating guidance.
Inspect the existing model before making requested changes. Onboarding itself
requires no model writes. For alternative plans, use a separate Scenario rather
than changing Main, following the product's current guidance and user intent.

Ask what the user wants to work on next if they have not already said. Account
access, MCP connection, and a working financial model are separate milestones;
report which are complete.

## References

- Product onboarding and operating instructions: https://docs.cfo.ai/agent-instructions.md
- MCP client setup: https://docs.cfo.ai/integrations/mcp-server.md
- Product documentation index: https://docs.cfo.ai/llms.txt
- Site and product overview: https://cfo.ai/llms.txt

If a client command or product workflow has changed, consult these current docs
and the client's official documentation before retrying.