Command line

The bandicot command is inside the app, at /Applications/Bandicot.app/Contents/MacOS/bandicot.

Call it by that full path, or make an alias for it. A bandicot already on your PATH may be an older build without rpc. bandicot version prints the build, and bandicot --help prints every command.

bandicot rpc <method> '<json params>' sends one call to the running Bandicot and prints its answer as JSON.

Bandicot must be open. Every method is listed by bandicot methods, with an effect: read looks, write changes Bandicot's state, spends starts model work, and outside reaches the network, a browser, the Mac or Team members. Anything but a read needs --yes, and so does reading a path outside the project.

A result that grants access prints its tokens redacted unless you pass --reveal. --events <s> keeps printing that connection's events for that many seconds after the answer.

--timeout and --events take seconds, and a decimal such as 2.5 is accepted. A number outside the range in the table, or a word, is a usage error (exit 2). If Bandicot is not up yet, rpc retries the connection for up to ten seconds before it gives up with exit 3.

1. 1Open Bandicot.
2. 2List what you can call: bandicot methods --pretty.
3. 3Read something: bandicot rpc session_list --pretty.
4. 4Change something only with --yes, for example bandicot rpc session_rename '{"session_id":"…","title":"Draft"}' --yes.
5. 5Read the state back with a read method to confirm the change.

bandicot rpc exits 0 when the app answered, 1 when it refused, and 2 to 7 for problems before or around the call.

After exit 4 to 7, a call that is not a read may already have run. The message names a read method to check with: read the state before you call it again.

bandicot history search <query> searches this folder's conversations for the active profile and prints JSON; it works with the app closed.

bandicot history read <session-id> reads one conversation and bandicot history sessions [query] lists them. The project is the folder you run it in. It never asks the running app, so it cannot change which chat is open.

Errors are one JSON line on stderr with a code: exit 1 means refused and exit 2 means a usage problem, including invalid_args.

bandicot setup plan FILE shows what a setup file would change, and bandicot setup apply FILE applies it after you consent.

In a terminal, apply asks for each key the file needs; pass --credential PROVIDER=stdin or env:NAME to supply one without typing it. Sensitive changes need --allow ID or --allow-sensitive after reviewing the plan. bandicot setup undo reverses the last apply, and bandicot setup export --out DIR writes your setup to a folder.

check validates a file, plan previews it, apply writes it, undo reverses an apply, export writes your setup out and receipts lists past applies.

check FILE validates the file alone and reads none of your settings. plan FILE prints each change with a mark (+ to create, ~ to change, - to remove) and flags the sensitive ones. apply FILE writes after consent. undo [RECEIPT] restores what an apply changed, the last one by default. export --out DIR writes a setup folder. receipts lists every apply on this Mac.

Every command takes --json for one JSON result instead of text. Keys are never read from arguments: a value for --credential must be stdin or env:NAME, so it never shows in a process list.

bandicot brain runs the agent's memory tool from Terminal on the folder you are in, or on notes kept everywhere with --global.

context <query> finds the few notes that fit a question, recall [query] lists or searches notes, remember <name> <text> saves one and forget <name> deletes one. Put --global last to use the notes kept everywhere; for recall it may come right after the command. A name or query must not start with a dash.

It exits 0 when it worked, 1 when the memory tool refused (the message says why, for example a note that is too large or no note by that name), and 2 for a usage mistake such as a missing name or an unknown option. See Memory for where the notes live.