Skip to main content
Assistant works with the scripts and running Rebind app on your computer. It can inspect files, make edits, execute inline Luau, and use the Rebind CLI. Rebind’s models, Fast and Smarter, are part of Rebind Pro and need you to be signed in. A model of your own works without signing in. This page covers the in-app Assistant. RebindGPT on the website drafts code but has no live access to your Rebind app or CLI. The rebind ai terminal agent has its own approvals and sessions.

Open Assistant

Open a file in Scripts, click Tools, then choose Assistant. Describe the task in the message box. You can attach images or text files for context. Opening another file changes the editor context without starting another chat. Use New chat to start a separate conversation, or the chat menu to return to an earlier one. Messages sent while a turn is working queue behind it.

Manual and Auto

Choose the mode in the message box before sending a request. In both modes, Assistant inspects and edits files and runs inline Luau without a separate confirmation. Inline Luau has the full SDK, including input, output, networking, and rb.system.exec. The modes differ in what Assistant asks about first:
  • Manual asks before it starts a saved script, runs a shell command or a CLI command outside the read-only list, or clicks, fills, or runs JavaScript on a page.
  • Auto does all of those without asking. A script it starts can capture your keyboard and mouse input as soon as it loads.
The mode cannot change while a turn is working.

Review actions

Expand tool entries to inspect their result or error. Inline Luau entries show the exact source sent to Rebind. File edits include a diff. A long-running CLI or shell command’s entry has controls to refresh its output and stop the process. A command that starts a Rebind script can leave that script loaded after the command ends; inspect it in Scripts or with rebind ps.

Show Assistant a task

To automate something you already do, ask Assistant to watch you first. It says what a recording captures and starts one when you agree. Do the task once, then press F8, choose Stop observation from the Rebind icon in the menu bar (macOS) or tray (Windows), or click Stop next to Observing in the status bar. While it records, the chat shows the recording with its time, its counts, and a Stop button, and a strip at the top of the screen shows the time and how to stop it. While it records, Rebind logs every key and button by name with its hold time, mouse movement, scrolling, and the active window, and takes screenshots (on macOS, when Screen capture is on in Settings). Passwords are recorded like any other keys, so do not type secrets during a session. The format is in rebind observe. Assistant reads the log and can look at screenshots from the recording. Before the first one goes to the model, it shows them and asks, in Manual and Auto mode alike. The model is Rebind’s AI service, or the server of a model you added. Sharing covers that recording and that model for the rest of the conversation; after you switch models, reload, or restart, it asks again. Chat history keeps the screenshots’ file names, not the images. With a model you added that has no image support, Assistant works from the log alone. Assistant then lists the steps it saw for you to confirm, writes a script, and may ask to watch again for a step the first session did not show. Recordings stay in Recordings until you delete them.

Research helpers

For a question that needs a lot of reading, Assistant can hand the research to a helper. The helper starts with only the task, reads files, logs, the documentation, and web pages, and writes a short report; what it read stays out of the chat, which keeps a long conversation usable. A helper cannot change files, run or stop scripts, run commands, or ask you anything. The helper works in the background, so you can keep chatting while it reads. Its row shows what it is reading, then its report. When the report is ready, Assistant reads it in a turn of its own and answers in the chat; if Assistant is busy with another message, the report waits until that turn finishes. One helper runs at a time. A helper takes at most 24 steps, and each one counts against your AI allowance like any other, as does the turn that reads its report. To end a helper early, click the stop button on its row.

Stop a turn

Click Stop in Assistant, or press Esc while its message box is focused.
  • Stop cancels the next tool call or model request, ends a research helper started in the same chat before its next step, asks running commands to exit, and drops messages queued behind the turn, including a helper report waiting for its turn. The report stays in the helper’s row, and Assistant can still read it there.
  • Anything already sent can still finish, and its result appears in the chat if you are still signed in to the same account. While Stopping shows, Assistant is waiting for that work to finish.
  • Stop does not undo edits or input that already happened.
To stop a loaded script, use its Stop control in Scripts. For input recovery, use the global kill switch.

History and recovery

Chat history, attachments, drafts, and question forms are stored locally in the app or browser, scoped to your signed-in account. They do not sync to another computer. Clearing that local storage removes them. Queued messages are not restored after reload. After a reload, an open question or a wait on a running process picks up where it left off, as long as you are still signed in. Your saved answer goes in before any follow-up you queued. A wait watches the process on your computer and makes no model requests while it waits. Restarting Rebind ends the process, so a wait on it cannot resume. A research helper still working at reload does not resume either: its row is marked interrupted, no report arrives, and the steps it had taken still count against your allowance. Other unfinished operations are marked interrupted. Check the affected files, scripts, or processes before retrying; an interrupted record does not prove that the operation never ran. If history cannot be saved, keep the window open and copy any messages you need before closing it.

Share a chat

Click Share chat at the top of Assistant, then Create link, to publish a copy of the chat on rebind.gg. Anyone with the link can read it; search engines are asked not to index the page. The copy holds your messages and the text files you attached. From Assistant it holds the replies, the reasoning, one line for each step, and the lines it changed in each file it wrote, with a few lines around each change. Long changes are shortened, and encrypted script content is hidden. What its tools returned, such as command output and files it read, is left out, though a reply or the reasoning may quote it. An image shows by name only. Messages sent later are not added. Rebind cannot delete a link once it is created.

Use your own model

The model menu in the message box lists Rebind’s models and any model you add. Choose Add your own model… at the bottom of the menu to connect one of these:
  • Ollama 0.12.6 or newer.
  • LM Studio, or any other server that speaks the OpenAI chat completions API, such as OpenAI or OpenRouter.
The Assistant model menu listing Fast and Smarter, each included with Pro, above Add your own model

The model menu: Rebind's models, and Add your own model at the bottom. Sample data.

Pick a provider or enter the server address, then an API key if the server needs one. Rebind lists the server’s models; tick the ones to show in the menu. Assistant edits files through tools, so a model must support tool calls and have a context of at least 32,768 tokens. Rebind checks both for Ollama. Other servers do not report tool support, so pick a chat model that has it. Your own models are free on every plan, work without signing in, and do not use your monthly AI allowance. Signed out, Assistant’s web fetches are limited to docs.rebind.gg. Model requests go from your computer straight to the server you added. Rebind’s API builds the prompt from the workspace file list, runtime state, your AGENTS.md rules, and the open file’s path, but not from the conversation. When an OpenAI-compatible server reports a large enough context for the model, requests also carry Rebind’s full documentation, and a provider that charges per token charges for those tokens on your key. Documentation search and web fetches go through Rebind’s API with the query or address the model chose, so your own models stop working while Rebind’s API is unreachable. The key is saved in Rebind’s settings file on this computer. On macOS, only the account that runs Rebind can read that file. It is left out of cloud backups and support logs, so servers you add are not restored on another computer. Shell commands and inline Luau can read the file. Of the two, only a shell command asks first, and only in Manual mode. Ollama’s default context is often too small for Assistant, so Rebind asks Ollama for a 32,768-token context on every request. That uses more memory than the default; on a smaller GPU the model runs partly on the CPU and more slowly. LM Studio sets the context when a model loads: load it with 32,768 tokens or more. When a chat no longer fits a model’s context, Assistant stops with an error that names the server. Start a new chat to continue. Small local models can make destructive edits, such as replacing a whole file when asked to add one line. Check the diff on each edit card and use its Revert button to undo a bad edit. To change or remove a server, open the model menu and choose Edit followed by the server’s address. Untick a model to hide it, or choose Remove server to delete the server and its key. If the model in use is removed, Assistant switches to Fast.