Provider in the model picker
Command Code appears as a provider. Pick a model per thread like any other harness.
ACP harness · MIT · verified on command‑code v1.89.1
The commandcode CLI already speaks the Agent Client Protocol. T3 Code already
speaks it too. This repo connects the two, so your Command Code subscription shows up as a
normal provider next to Claude, Codex and Cursor — no proxy, no wrapper, no API key.
$ git clone https://github.com/jovylle/t3-command-code-harness.git
$ cd t3-command-code-harness
$ ./install.sh
==> commandcode: /Users/you/.local/bin/commandcode
==> backup: ~/.t3/userdata/settings.json.bak.2026-11-16T10-25-41-000Z
==> added providerInstances.acpRegistry_command_code
==> ready - 98 models
==> first: deepseek/deepseek-v4.1-flash, ...
Done. In T3 Code open the model picker and choose Command Code.
Nothing runs in the background. No port is forwarded. T3 spawns the CLI, sends
initialize and session/new, and every turn after that is plain ACP traffic.
What you get
Command Code is exposed the same way every other T3 provider is — which means threads, model switching, permission modes and usage tracking all behave normally.
Command Code appears as a provider. Pick a model per thread like any other harness.
The list comes from Command Code, not this repo — roughly 98 models today, default deepseek/deepseek-v4.1-flash.
It reads ~/.commandcode/auth.json. If you are signed in to the CLI, T3 just works — no re-auth, no token copy-paste.
T3's mode switches land on Command Code's own modes: Standard, Accept Edits, Plan, Don't Ask, Bypass.
Install
The script is idempotent: run it twice and it tells you there is nothing to change. Your settings file is backed up before every write.
T3 Code (with the acpRegistry provider driver), the CLI, and a Node runtime — T3 ships one.
$ commandcode --version
$ commandcode login # only if you are not signed in
$ ls ~/.commandcode/auth.json
It finds the binary (or honors COMMANDCODE_BIN), edits
~/.t3/userdata/settings.json, and waits for T3 to hot-reload the provider.
$ git clone https://github.com/jovylle/t3-command-code-harness.git
$ cd t3-command-code-harness
$ ./install.sh
Open T3, hit the model picker, choose Command Code. Restart T3 only if the provider does not appear.
Add this under providerInstances in ~/.t3/userdata/settings.json, with your own command -v commandcode path:
{
"providerInstances": {
"acpRegistry_command_code": {
"driver": "acpRegistry",
"displayName": "Command Code",
"enabled": true,
"config": {
"source": "local",
"commandPath": "/Users/you/.local/bin/commandcode",
"commandArgs": ["acp"]
}
}
}
}
Removes the provider entry (backing up settings first) and deletes the probe cache. Restart T3 if it is still listed.
$ ./uninstall.sh
Gotchas
Learned the hard way, written down so you do not repeat it.
In a non-prompting permission mode, any tool outside the allowlist is denied with no prompt.
It looks like the agent is broken when it is really the harness refusing. Use Accept Edits or Plan,
or add explicit permissions.allow rules.
If the CLI updates while a session is live, that session can break. Finish or reopen the thread instead of expecting the old one to keep going.
T3 asks the CLI. Different Command Code plans expose different models, so two people can see different pickers on the same commit.
Verified against v1.89.1. If acp disappears from commandcode --help,
the flag was renamed — the fix is one string in settings.json.
One script, one settings key, your existing login.