Agent mode

In agent mode the AI works on the project in the current folder: it reads files, searches, edits, and runs commands until the task is done.

atlas agent "add unit tests for src/date.ts"
atlas agent -p groq "fix the failing lint errors"

In the chat:

/agent rename the config option "verbose" to "debug" everywhere

From the chat, the agent sees the conversation so far. Agent mode needs an interactive terminal because it asks for approvals.

Tools

Tool	What it does	Asks first
read_file	Read a text file (up to 2000 lines per call)	no
list_dir	List a folder	no
search	Regex search across files (skips node_modules, .git, dist)	no
write_file	Create or overwrite a file	yes
edit_file	Replace exact text in a file	yes
run_command	Run a shell command in the project folder	yes
MCP tools	Tools from your MCP servers	yes, always

Approvals

Before each change you see a preview: a diff for edits, the full command line for run_command. Then pick:

Answer	Effect
Yes	Run this one call
Yes, don’t ask again this session	Run it, and allow this tool without asking until the agent finishes
No, do something else	Skip the call; the AI is told you declined

↑ ↓ choose, Enter confirms, Esc means No. Esc while the agent is working stops it.

At the end you get a summary: steps taken and files changed. The agent stops after 25 steps; ask it to continue if needed.

Safety

• Project only. File tools resolve every path inside the folder where you started the agent. Paths outside it, including through symlinks, are refused.
• Secret files are blocked. The agent cannot read or write .env and .env.*, *.pem, *.key, *.p12, *.pfx, SSH keys (id_rsa, id_ed25519 and similar), anything under .ssh/ or .aws/, ~/.the-atlas/config.json and ~/.claude/.credentials.json.
• Commands. run_command uses Windows PowerShell 5.1 on Windows and bash elsewhere. It has a 2 minute timeout by default (10 minutes at most) and is killed when you stop the agent. Output is truncated to keep the context small.

Caution

run_command and MCP tools can do anything your user account can. The path rules above apply to the built-in file tools, not to shell commands. Read each command before you approve it, and prefer Yes over Yes, don’t ask again for run_command. Use git so you can review and undo changes.

Choosing a model

Agent mode needs a model that supports tool calling. In auto mode, if a model rejects tools, The Atlas moves on to the next one. To pin a model:

atlas agent -p openrouter -m openrouter/free "..."