Skip to main content
A chat agent is not a chatbot with a text box. It is a process with tools. ADE gives every chat the same core set, then adds what the session’s runtime and role allow. There are four layers. Most questions about “what can an agent do?” are really questions about which layer is in play.

​
1. Universal tools

Available to every agent — regular chat, personal chat, and the CTO — in every provider. These are real tool calls, not prompt instructions. Each one produces an event in the transcript, which is why you can inspect every edit and command after the fact.

​
2. ADE surfaces through the CLI

The agent’s shell has the ade command. That is how a chat reaches ADE itself:
You do not install this separately: the same binary that runs the machine is on the agent’s path, and the agent’s system prompt explains how to discover the rest with ade help.
This is why “create a lane”, “open a PR”, and “capture proof” are not separate tool rows. They are commands, so they compose, work from any provider, and stay identical to what you would type.

​
3. Provider-native tools

Each runtime also brings its own tools, and ADE preserves them rather than flattening them into the universal set:
  • Claude Code brings its native AskUserQuestion tool, its own subagent tool, and its own web tools.
  • Codex brings its web, MCP/connector, and image capabilities.
  • Cursor, Droid, OpenCode, and the ACP providers bring their own tool loops and MCP support.
Provider-native activity is normalized into compact transcript events, so every client can show what happened without dumping raw SDK envelopes.

​
4. CTO operator tools

The CTO is the one agent with elevated control-plane tools: spawning chats, creating lanes, reading and writing Git and PR state, running tests, reading and updating Linear, and reading and writing its durable memory. Regular chats never see these.

CTO

The project-level operator and its elevated tool set.

​
Permissions decide what runs

Tools exist; the permission mode decides which ones execute. The mode is per session and visible in the composer. full-auto does not hide what happened: writes and commands still emit approval events for review.

​
Seeing the toolset

Open the tools panel from the composer to see what the current session can reach and which actions will require approval.

Chat capabilities

The task-level view: what a chat can do end to end.

Permissions

Modes, approval gates, and what they protect.