1# fx
2
3T3 Code can run [fx](https://fx.sh). fx uses the login on the machine that runs the T3 Code server.
4
5## Set up fx
6
7Install fx on that machine:
8
10 curl -fsSL https://fx.sh/setup.sh | bash
12
13Run `fx login` in a terminal on that machine.
14Open **Settings → Providers**, enable fx, and refresh the provider.
15
16If `fx` is not on the server's `PATH`, set **Binary path** to the executable.
17
18Every fx instance on the same machine shares that machine's fx login. T3 Code does not keep a separate login per instance.
19
20## Models and effort
21
22The model picker lists **Default** first. That option runs the model `fx status` reports on the server's machine. A model set only for one project in fx's settings is not picked up, so choose it from the list instead. If the thread's fx backend does not offer that model, the thread keeps its current one. The other entries come from `fx models`. A custom model you add in the provider settings goes to fx as is, and fx reports an error if its backend does not offer it.
23
24fx also has an **Effort** control. A new thread starts at Medium.
25
26T3 Code does not show fx's backend selector. To use Codex, Grok, or Vercel AI Gateway, run `fx provider` on the environment's machine.
27
28Type `$` in the composer to list the skills fx finds in its skill folders, in the order fx's own skill menu shows them. When two different skills share a name, the menu lists the one fx loads for that name: the first copy in fx's folder order, starting with the workspace. The menu shows one skill per name regardless of case; to run a skill whose name differs only in case, such as `Review` next to `review`, type its exact name.
29
30## Permission modes
31
32fx offers two composer modes.
33
34| Mode | Behavior |
35| -------------- | ----------------------------------------------------------------------------------- |
36| **Supervised** | fx asks before it changes files or runs commands. Read requests run without asking. |
37| **Auto** | fx handles routine actions itself. Anything it still asks about reaches you. |
38
39T3 Code does not offer **Full access** or **Auto-accept edits** for fx. A thread already set to one of those runs in **Supervised**. Changing the permission mode restarts the connection and reloads the same fx session.
40
41On fx versions after [vercel-labs/fx#1189](https://github.com/vercel-labs/fx/pull/1189), fx saves the mode T3 Code sets as its default permission mode. T3 Code sets it each time it opens an fx thread, so fx in the terminal then starts in that thread's mode.
42
43A message steered into a running turn stops that turn, and fx answers the new message in the same session.
44
45## Sub-agents
46
47When fx hands work to a sub-agent, it opens in its own thread under the parent. That thread shows the sub-agent's task and its reply, which fx shortens to a couple of lines. fx doesn't share the sub-agent's individual steps.
48
49fx doesn't pass on a sub-agent's permission requests, so a sub-agent that needs one waits until you press **Stop**. This mostly happens in **Supervised**. In **Auto**, sub-agents can finish routine work by themselves.
50
51## Not supported yet
52
53**Full access**.
54**Auto-accept edits**.
55Commit messages, pull request text, branch names, and thread titles. Pick another provider for those.
56
57See [provider setup](./install.md#providers) and [Permission modes](./permission-modes.md).