Kaitoi Code.
Penny in your terminal. Code works with local files in a folder you trust, Kaitoi projects and graphs, runtime recipes, and approved local commands, and can hand long coding tasks to the Crew.
Kaitoi Code is Penny in your terminal. It can use Kaitoi Core API tools, inspect and run node graphs, work with local workspace files once you trust a folder, run local runtime recipes, and download or upload files with your confirmation.
Start an interactive session:
kaitoi code
In an interactive session Code offers to use the
current folder as its trusted workspace, so --workspace . is normally
unnecessary. To target a different folder:
kaitoi code --workspace /path/to/project
Ask a one-shot question instead:
kaitoi code "Create a project with a text-to-image node and tell me what you did"
Slash commands
/login refresh your Kaitoi login token
/status show auth, model, workspace, and mode
/providers list available providers and models
/provider gemini switch model provider
/model <model-id> switch model
/workspace <path> set or show the trusted workspace
/mode plan switch to read-only planning
/commands on enable per-command confirmed local shell commands
/permissions show session approvals
/doctor show local Code diagnostics
/config init create a persistent Code config file
/provider, /model and /commands change the current session only. They do
not silently rewrite your config file.
File access is explicit on purpose
Code only reads or writes files inside a trusted workspace, and every write, download and upload asks first. This is a deliberate constraint rather than an oversight: an agent that can quietly write anywhere on your machine is not one you can leave running.
Handing a task to the Crew
For a larger coding task, ask Code to use the Crew. Code inspects the trusted Git workspace, gathers any missing choices, and shows the exact proposed scope, models, verification commands and budgets before asking for approval.
Once approved, the run is durable: it can plan, implement, test and review in the background even after the Code session ends. Afterwards, ask Code in plain language to check status or progress, read the result and its evidence, or propose a control action such as steering, pausing, resuming, interrupting or cancelling.
Starting a run and changing its controls require explicit approval. Status, progress, evidence and results are read-only and do not.
The Crew is available to Code as a permission-aware tool, not as a
slash command. Advanced users can inspect the direct surface with
kaitoi crew --help.
Modes
Code has three modes. plan is read-only. build edits files in the trusted
workspace. crew keeps workspace reading, the Crew's own tools
and the Kaitoi platform tools, but has no file-writing, command, install or
delete tools at all, so a request to build something becomes one exact start
proposal for your approval rather than files changed in the chat session.
kaitoi code --mode crew
Inside a session, /mode crew switches, and /crew <task>
switches and sends the task in one step. A tool outside the current mode is
refused rather than run, so the guarantee does not depend on the model choosing
well.
Current limitations
- Local shell commands are off until you enable them with
/commands on, and each one is confirmed individually.
Last reviewed against Kaitoi on 16 September 2026. View this page as Markdown