AI Assist And Dictation

AI Assist hands short writing jobs to your selected provider, including connected ChatGPT accounts and agent CLIs. Dictation turns speech into text in the fields where you write prompts. Neither needs an Alera account.

What AI Assist Writes#

• Commit messages from your staged changes: Generate commit message with AI in Source Control.
• Pull request titles and descriptions: Generate Title And Description With AI in the pull request form, once a base branch is selected.
• Workspace names and branches when you create a workspace From Prompt (see New Workspace), plus a section when Auto Assign Section is on. Hand Off suggests a branch name the same way.
• Agent titles: new agent tabs are named from their first prompt while Auto-Generate Agent Titles is on. Right-click a tab for Generate Title or Regenerate Title.
• Reading diffs: Generate Reading Diff in a diff tab writes a behavioral overview and a condensed diff. It asks first, because it spends quota, and it is not a code review.
• Speech messages: optional cleanup or summary of dictated text.

Commit messages and pull request details land in the form for you to edit, and are dropped if you change the field while the job runs.

Choosing The Agent#

Configure it in Settings → AI Assist. Enable AI Assist is on by default, and the Agent is Codex, which runs codex exec in a read-only sandbox. You can pick ChatGPT, Claude Code, GitHub Copilot, Cursor, Antigravity, OpenCode, OpenCode 2, OpenCode Go, Pi, Amp, Grok Build, fx, or Custom Command, then a Model and, where the model supports it, a Reasoning level. Refresh Models fetches the current list.

Each job type (Commit Messages, Pull Request Details, Reading Diffs, Workspace Identity, Agent Titles, Speech Messages) has its own group where you can keep the Global agent and model or override them, and add Instructions such as your commit convention.

A Custom Command receives the prompt on stdin, or as an argument wherever you write {prompt}, for example llm --system commit-message. It cannot process speech.

CLI providers must be on your PATH and signed in. If one is missing, the job names the command it looked for. Jobs use the configured timeout, two minutes by default.

Using Your ChatGPT Plan#

Under Settings → AI Assist → ChatGPT Account, choose Continue with ChatGPT and authorize Alera in your browser. Then choose ChatGPT as the global agent or for specific job types. No API key or Codex installation is required. Available models come from the selected ChatGPT account. Saved Accounts keeps account and workspace registrations separate; Use Account switches between connected registrations, and Sign Out clears the selected account’s tokens.

Eligible requests use your ChatGPT plan or credits you authorize in ChatGPT. Manage Usage opens ChatGPT’s app limits and access settings. Reaching a limit stops the request; Alera does not silently switch to API billing. Credentials stay in the runtime’s protected storage and never enter synced settings or the phone. Each remote runtime requires its own connection.

This integration generates text, including cleanup of an existing transcript. OpenAI’s Sign in with ChatGPT preview does not support audio input or transcription, so choose a separate transcription engine for dictation.

Settings → Text Actions uses the same agents: save an instruction with New Action, then apply it to selected text from the Text Actions entry of the right-click menu.

Dictation#

Turn on Enable AI Dictation in Settings → AI Dictation; it is off by default. Then use the microphone button in the Terminal Composer, an agent profile’s starting prompt, From Prompt, the commit message, or the pull request title and description. Click Start Dictation, speak, and click Stop Dictation to insert the text.

Pick a Transcription Engine:

• Local Whisper (the default) keeps audio on this computer. Download a model under Local Whisper Models and choose Use Model: Whisper Tiny, Whisper Base (recommended), Whisper Small, or Whisper Large V3 Turbo Q5_0. Downloads come from the whisper.cpp model repository, run one at a time, resume after an interruption, and are checked against a pinned SHA-256 checksum before use.
• Codex Subscription (Experimental) uses the realtime API included with a Codex subscription, through your signed-in codex CLI. Alera never reads your Codex credentials. Realtime Model is an optional override.
• OpenAI-Compatible API sends the recording to any speech-to-text endpoint: set the Base URL (default https://api.openai.com/v1) and Model (default gpt-4o-mini-transcribe). The optional API Token goes to the system credential store (on Linux, a private file only you can read, if no keyring is available), is tied to that API’s origin, and never enters Settings. Local servers that need no token work too.
• System On-Device on macOS, offered only when the system recognizer can guarantee offline processing, and System Recognition on Windows, which may send audio to Microsoft and needs Allow Online Speech Recognition.

Both remote engines need Allow Remote Audio Processing, and the recording is deleted locally after transcription. Request Timeout allows 5 to 300 seconds, 60 by default. Leave Language blank to detect it automatically.

Speech Processing → Automatic Processing can Clean Up or Summarize the transcript with the agent set for Speech Messages in AI Assist. If that agent fails, the raw text is inserted.

Test a configuration in the Test AI Dictation field. Runtime Update Required means the running runtime is older than the app; replace it with Update Runtime (see Troubleshooting).

On The Phone#

The Android app has its own AI Dictation screen in Settings. Its Processing Location is This Device, with Whisper models, the system recognizers, and an OpenAI-compatible API whose token stays in the phone’s secure storage, or Paired Device, which hands the recording to the computer’s Whisper model, API token, or Codex subscription. Except with the system recognizers, you can play a recording back before choosing Transcribe or Remove Recording. See Mobile Companion.