bandicot
Bandicot Docs
Chapter IStart · Available now

Best practices

Short habits for working with Bandicot agents: clear tasks, plan first, Versions on, secrets in the vault, cost in view and a backup of your history.

Available now. A current user-facing path exists within the stated scope.

Write the task so you can tell when it is done

Name the files, say what you will notice when it works, and state what must not change.

A task that says Make the sign-up page better leaves the agent guessing. A task that names the file, the result and the limit does not: In @signup.html, show the error under the email field, and leave the layout alone.

If you are not sure what you want, ask for /grill first. It asks one question at a time about what you will notice and never about technology.

While the agent works, press Enter to queue a follow-up for when it finishes, or ⌥Enter to steer the work in progress. Press ⌘. to stop a run.

Plan first when a change touches many files

Use /plan for a change that has more than a few steps, read the plan, and edit it before the agent starts.

/plan researches and writes a plan where each step ends with a check, and it does not start the work. /super-plan writes a workspace plan and a plan for each part. Reading ten lines is cheaper than reviewing a wrong change.

Edit the plan in the Plan pane: add, reorder or drop steps. Use New plan to queue the next piece of work. A queued plan is seen only once the plan above it is finished. See The plan.

Keep projects small and named

Make one project for each piece of work, name it for the result, and put the lasting rules in the folder's handbook.

A project or workspace keeps its own agents, plan, handbook and version history, so a small one is quicker to read and to restore. Each agent starts from the folder's AGENTS.md, or its CLAUDE.md when there is no AGENTS.md. Put rules that should last there, not in a message.

Give an agent a name and a hotkey with Session profile so you can find it later. When the agent learns something worth keeping, a notice asks Remember this for next time?. Press Keep and it becomes a note. See Memory.

Do not use your home folder as a workspace if you want history. Versions does not keep it, so open a folder inside it.

Turn on Versions before the first agent run

Turn on Versions for a folder before an agent first changes it, because only changes made after that are recorded.

Agents keep working in a folder without Versions, but their changes are not recorded there. With it on, you can put back one file, the whole folder or everything one run changed. Before a risky change, select the current version and choose Name this point.

Remember what Versions cannot reverse: shell commands, web requests, outside tools and files written outside the folder. For those, read the card before you allow them. See Undo everything an agent run changed.

Answer cards narrowly and write `ask` rules for what matters

Allow once or for this agent first, allow a whole workspace only after you have seen the same kind of call several times, and add an ask rule for anything you always want to see.

Bandicot has one mode, Auto. Ordinary work inside the workspace runs without a question, and four kinds of risky call are checked first. What you answer is narrower than it looks: an allow covers calls of the same kind, not every call of that tool, and it never covers a dangerous shell line.

For work that must always stop, write a rule with the action ask in permissions.json; it is the one way to get a Needs your call card for a call Auto would otherwise judge by itself. On the Decisions page, the Risk check row says that when it is off, you are asked as usual. A company can set Always ask before acting for every Mac.

A card nobody answers within sixty seconds counts as Deny. See Approvals.

Use Decisions and Checks, and let agents read both

Give Decisions a model so quick questions do not use your main model, and set up Checks so an agent can prove its work.

Decisions is a small, fast model that Bandicot asks quick questions while an agent works, such as whether a tool call is clearly safe. Each job has a switch in AI Gateway › Decisions, and a switch takes effect on the next question. A Laya model runs on your Mac and costs nothing per ask. See Choose the Decisions model.

Checks runs your tests, lint and build on your own Mac in a clean copy of a saved version and keeps a signed result. An agent can start the checks, wait for them and read each pass or fail, so it never asks you to run them. See How agents run checks.

Watch the context meter and the usage pane

Check the Context usage meter during long work and the Agent cost figure after each run, and compact before the window fills.

The meter on the Context page shows how much of the model's window the agent fills, in slices such as Conversation and Tool definitions. When it says Compaction due, the agent is near the limit. Type /compact to compact now, or set Compact at in the Compaction card under Decisions.

Cost figures are estimates, not invoices. A cost marked Estimated means the provider did not report its price. For a total across profiles, open AI Gateway › Usage and choose the range. See Usage and cost.

To cap what Bandicot spends on its own, set Automatic AI allowance · USD and press Save account limit.

Keep secrets in the vault, never in a prompt

Use Share a secret or the Passwords vault for every key and password, so the value never reaches the model.

Keys, sign-ins and recovery keys sealed in Settings › Passwords can be used by an agent where you allow, and the agent never sees the value. A kept secret can be read only inside a command, for twenty-four hours.

If a page needs a password and you lent a Chrome tab, type it yourself when the Bandicot needs you card appears. Forget any secret you no longer need with Forget. See Secrets and passwords.

Give parts to subagents, and keep `Start without asking` off at first

Hand independent parts of a big job to subagents, and approve each start until you trust the pattern.

A subagent runs in its own agent, so its reading and trial runs stay out of the parent's context. Choose Start on each card to begin with. Start without asking lets the agent start the next ones on its own, and it is the Start subagents without asking switch under Agent › Capabilities.

For a helper you use often, save a markdown file in .bandicot/agents with a description and your instructions, and ask the agent to hand the job to it by name. A subagent runs on the model you chose under Helpers, or else on the parent's model. See Your own subagents.

Use automations for work that repeats, and start with `Ask me first`

Turn a repeating job into an automaton, point it at what matters, and let it ask before it prompts an agent until you trust it.

An automaton wakes only on what you pointed it at, and it waits for quiet before it wakes, so it does not run on every event. Each trigger says what to do: Just tell me, Prompt Bandicot or Ask me first. Prompt Bandicot sends a prompt up to six times an hour, and past that it asks you first.

Read Last report and the dots under Runs regularly. If you have an Anthropic or OpenAI API key, Slow lane for automations runs the work at half price. See Automations.

Use a profile for each person or role

Keep work and home, or each family member, in separate profiles so memory and agents do not mix.

A profile keeps its own agents, memory and automations apart from every other profile. Free has 1 and Pro has 5, with more at $1.99 each. Family links 6 accounts, and Business gives each person 5.

Use Clean this profile to delete a profile's agents when you finish a role, and export a profile as a zip when you move it. For a team, one setup file starts everyone the same way. See Profiles.

Match the model to the job

Use a stronger model for hard reasoning and a smaller or local one for simple, repeated work.

Set the model for each agent from the model menu in the prompt bar, and set Effort only where the model can think longer. The levels are None, Low, Medium, High, Extra high and Max, and the default for new agents is under AI Gateway › Agent › Behavior.

Put a model on each task with the Rack. Compare real asks before you switch with Compare models on these jobs. See Connect a model and Local models.

Back up the history and keep the recovery kit

Make a recovery kit, start second copies and keep the kit somewhere other than the Mac.

Versions › Backups keeps an encrypted second copy of every history in iCloud Drive or on a disk you plug in, and only your recovery kit opens it. The copy updates about every fifteen minutes. Bandicot does not keep the recovery words and cannot show them again, so write them down when you make the kit.

For a plain folder you can open anywhere, use Export a recovery copy… on the Storage page. See Keep a recovery kit and A second copy with your recovery kit.

Questions

What is the single most useful habit?

Turn on Versions before the first run. Then a wrong run is one click to put back, and every other habit on this page becomes less urgent.

Should I allow a whole workspace to save time?

Not at first. Allow once and Allow this agent are narrower, while Allow this workspace covers calls like this one from now on. A dangerous shell line is checked again even after an allow, so wide allows buy less than they seem to.

How do I keep costs down without losing quality?

Watch Agent cost after each run, use a smaller or local model for repeated simple work, and let Decisions answer quick questions on a model on your Mac. See How to save cost for the steps.

Where do I read what an agent actually did?

Open the agent's rows, the Timeline in Efficiency titled Everything this agent did, and the Diffs pane, which shows each file before and after.