Agents (MCP and the CLI)

An agent is an AI assistant that does tasks for you, such as Claude Code or Codex. An agent can use tabletop through MCP (Model Context Protocol). It can read your workflows, add and link nodes, start runs and read the results.

The CLI (command-line interface) gives you the same actions as commands in a terminal.

What an agent can do

Scope	What it lets the agent do
Read workflows	Open workflows, read the activity of the org, and read runs, results and versions.
Change workflows	Make, rename, copy, import and edit workflows. Move them to the trash. Pick a version of a node.
Start runs	Start, approve and cancel runs. Get the token estimate of a run.
Read billing	Read the token balance of the org.

A run that an agent starts costs tokens, as a run that you start. See the full list of tools in MCP tools.

Connect an agent with an API token

An API token is a secret key that lets an agent or a script act for you in one org.

Click your avatar, then Organization settings.
Open API tokens.
Type a Token name, such as “Claude Code on my laptop”.
In Days, type how many days the token works. Leave it empty for a token that does not end.
Select the scopes that the agent needs.
Click Create token.

The API tokens tab with Token name, Days, the scopes and Create token

features/agents-create-token

tabletop shows the token one time. It also shows a command for Claude Code and a command for Codex:

Click Copy next to the command of your agent.
Paste the command in a terminal and run it.
Click Done.

The new token with the Claude Code and Codex commands and the Copy buttons

features/agents-token-commands

note

Codex reads the token from TABLETOP_TOKEN. Set it in your shell profile, so that it is there in each new terminal.

Copy the token now. You cannot see it again. If you lose it, make a new token.

An agent that supports MCP sign-in can connect without a token.

Add tabletop to the agent with the MCP address https://app.tt3d.ai/mcp.
The agent opens tabletop in your browser. The page says that the agent “wants to use tabletop for you” and lists what it can do.
In In the organization, pick the org.
Click Allow.

The page Claude Code wants to use tabletop for you, with the org list, Deny and Allow

features/agents-allow

See all the steps in Connect Claude Code or Codex.

See the agent at work

When an agent edits a workflow that you have open, it shows as an avatar at the top right. The tooltip says “Agent of” and your name. A label moves to each node that the agent changes. See Live editing with your team.

Manage your tokens

The API tokens tab lists your tokens. Owners and Admins see each token of the org.

Column	Meaning
Name	The name of the token. “Service account” shows under the tokens of a service account.
Scopes	What the token lets an agent do.
Last used	When the token was last used, or “Never”.
Ends	When the token stops working, or “Never”.

To stop a token, click Revoke. Agents and scripts that use it can no longer use tabletop.

The list of API tokens with Name, Scopes, Last used, Ends and Revoke

features/agents-tokens

Use the CLI

The CLI is the tabletop command. Each MCP tool is also a CLI command, such as tabletop runs start.

Run tabletop login. Sign in in your browser.
Run tabletop org use with the id of your org.
Run the commands that you need.

For scripts, give an API token with --token or the TABLETOP_TOKEN environment variable.

note

A download of the CLI is not available yet. To ask about it, see Get help.

See every command in CLI commands.

Agent nodes

Some nodes can also use an agent as a provider: Agent: Claude or Agent: OpenAI. The agent runs in tabletop and makes the result of the node, such as an image, a model or a file. Write what the agent must do in Instructions. See Providers.