Beta AI assistance

Editora has optional AI, off by default and yours to configure. It comes in two parts: quick one-shot actions, and a full embedded agent.

Turning it on

1. Turn on Enable AI in Settings → AI (or view.toggleAiEnabled). This master switch gates everything on this page: while it is off, every AI feature is disabled whatever its own setting says.
2. Turn on the part you want: AI Actions (view.toggleAi), the AI Agent (view.toggleAgent), or both.
3. For AI Actions, choose a provider, give it a key or an endpoint, and press Test connection on the AI Actions page (or run AI: Test Connection, ai.testConnection). For the agent, install an ACP agent and check its command under Settings → AI Agent.
4. Inline completion has its own switch under AI Actions (view.toggleAiCompletion) and stays off until you turn it on.

AI actions

These call the model directly (streamed), enabled under Settings → AI Actions:

Rewrite Selection and Generate Commit Message put the model’s text into your work, so they apply a reply only when it is complete. A reply that was cut off at the model’s output limit, ended because the connection closed, came back empty, or was declined is discarded, and the status bar says which. A rewrite is also discarded if the selected text changed while the reply was being generated. Explain writes to a buffer of its own, so a partial explanation is kept and marked as incomplete.

Rewrite refuses a selection longer than 120,000 characters instead of sending part of it, because the reply replaces the whole selection.

The top of the AI Actions page has a status pill that shows whether the provider answers: a green Connected, or a red Not working with the reason. Opening Settings sends no request. Editora checks only while the AI Actions page is shown, with one tiny request, and shows a result from the last three minutes again instead of asking twice. Test connection beside the pill always checks again. While Enable AI or AI Actions is off, the pill is grey and names the switch that is off, the page’s rows are greyed, and a note says which switch to turn on.

There’s also inline completion: after a typing pause, a muted one-line ghost suggestion appears at the caret, and Tab accepts it. It uses its own fast model (default claude-haiku-4-5), separate from the action model (default claude-opus-4-8).

AI Agent

The AI Agent is a chat with an embedded coding agent over the Agent Client Protocol (ACP). The default is Claude Code’s claude-code-acp adapter, but any ACP agent works (Settings → AI Agent). The agent is a user-installed external tool, never bundled. Each client has a card with its command and a line saying whether it is there, for example claude-code-acp: found (/usr/local/bin/claude-code-acp) or gemini: not found; the card of the client in use is marked In use.

• Its file reads see your open buffers’ unsaved text.
• Its edits to open files apply as undoable buffer edits that you review and save.
• A write to a file that is not open goes to disk. The file’s previous text is saved to local file history first, and the write is refused if that copy cannot be made. The file keeps its encoding, byte-order mark and line endings.
• It cannot replace text it has not seen. A write is refused when the file or buffer changed since the agent last read it, including text you typed in the meantime; the agent is told to read the file again.
• Each action that needs permission pops a dialog. The dialog opens with the focus on the rejecting choice, has no default button, and ignores key presses for a moment after it appears, so a keystroke meant for the editor cannot approve a request.
• It works in the project open in its own window: that folder is where it runs, and its file reads and writes are confined to it. Without a project, the folder is that of the active file, or else your home folder. It can never touch Editora’s own configuration, or version-control metadata such as .git/.
• In a large file that has no undo, an agent edit is preceded by a Local History copy, or refused when none can be taken.
• Replies appear as they stream. Remote images in a reply are not loaded; the alt text and URL are shown instead.
• The transcript is bounded: one entry holds up to 200,000 characters (a longer tool result keeps its start and says how much was cut), and the whole transcript up to 1,000,000 characters or 2,000 entries, the oldest going first.

Commands: tool.agent (the tool window), agent.newSession, agent.resumeSession, agent.selectClient, agent.selectMode, agent.selectModel, agent.stop. AI is gated behind a master Enable AI switch (off by default) in Settings → AI.

Providers

Pick a provider in Settings:

• Anthropic API: the API key comes from the ANTHROPIC_API_KEY environment variable or a Settings override; models are configurable.
• Local (OpenAI-compatible): point every AI feature at LM Studio, Ollama, or any OpenAI-compatible server. It has its own endpoint, action model and inline model, and an optional key for a server that asks for one.
• LM Studio / Bionic: a separate local provider with its own endpoint, action model, inline model, and optional token. It can also supply the model for the OpenCode ACP agent preset without writing OpenCode config files.
• Codex: use the user-installed codex-acp adapter and an existing Codex login for Explain, Rewrite, and commit-message generation. These actions use separate text-only sessions and leave your Agent chat intact. Codex has its own Model field (blank uses the Codex default). Inline completion is unavailable with this provider, and it takes no API key or endpoint.

When AI Actions are enabled and connected, the editor right-click menu includes AI Actions → Explain and Rewrite. Generated explanation.md buffers show the provider and response model when the server reports it.

Each provider has its own stored key, endpoint and models: the Settings fields show only the selected provider’s, so switching provider never sends one provider’s key to another provider’s server. The ANTHROPIC_API_KEY fallback applies only to the Anthropic provider. And Editora won’t attach a key to a plain-http:// endpoint on another machine, refusing before it connects and asking you to use https or a loopback address, so your key never crosses the network in the clear. Plain-http on 127.0.0.1 (the usual local-inference path) is unaffected.

Where API keys are kept

A key typed into Settings, for any provider, is kept in the system keychain: the macOS Keychain, the Windows Credential Manager, or the Secret Service on Linux (GNOME Keyring or KWallet, through libsecret). Keys that an earlier version saved in settings.json move to the keychain at the first start; each one leaves the file only after the keychain has stored it and given back the same text. The note under API key on the AI Actions page says where the keys are.

Where no keychain answers (libsecret is not installed, or no Secret Service runs, as on a headless machine or in an SSH session), keys are saved in settings.json in your config folder as plain text, and the note says so. If the keychain that holds your keys does not answer at a start, a key typed in that session goes to settings.json, and the note says that too. If the keychain refuses a key, Editora puts the keys back in settings.json for the rest of the session. On a shared machine without a keychain, leave the Anthropic field empty and set ANTHROPIC_API_KEY in your environment instead.

Keys in the keychain are never exported. Export Configuration… leaves keys kept in settings.json out too, unless you tick Include AI API keys, and Reset to Defaults keeps your keys; see Export and reset.

What leaves your machine

AI Actions send the following to the provider you selected: Anthropic’s API, the endpoint you configured for a local or LM Studio provider, or the codex-acp adapter running on your machine for Codex.

• Generate commit message sends the staged diff (git diff --cached).
• Explain sends the selected text and the file’s language name.
• Rewrite sends the selected text, the language name, and your instruction.
• A staged diff, or a selection sent to Explain, longer than 120,000 characters is cut at that length. Rewrite does not send a selection over that length at all.
• Inline completion sends the language name and the text around the caret, up to 4,000 characters before it and 1,000 after, once per typing pause. It only asks when the caret is at the end of a line.

The AI Agent gets your prompt. With Include editor context in prompts on (the default; view.toggleAgentContext), each prompt is also prefixed with the active file’s path, the cursor line, and the first 200 characters of any selection. Anything else the agent reads, it requests itself, within the limits listed under AI Agent.

Related

For letting an external agent drive Editora (rather than Editora calling a model), see the MCP server.