Agents, models and modes

Agents

Raven runs two coding agents: Claude Code and Codex. Both ship inside the app, so there’s nothing to install — connect your account in Settings › Account › Agents and the app runs the bundled binary on your behalf.

They’re independent. You can connect one and not the other, and different workspaces (or different tabs in the same workspace) can use different agents. Whichever you pick, the agent runs locally against the workspace’s checkout with the same file, terminal, and tooling access.

Saved accounts

Both agents support multiple saved accounts, and Raven runs turns only on accounts saved here — never on a login the CLI happens to have elsewhere on your machine. If the active agent has no account, sending shows a card in the chat with a button to set one up.

Choose Add account and sign in: Claude Code through the browser, Codex through a ChatGPT account or an OpenAI API key. Optionally name the account (Personal, Work), then Save account to keep it for later or Save and use to make it active.

Each agent has one active account across all chats. Use account switches that agent’s next turns to the selected account, including turns in existing conversations; turns already running finish on their original account. Switching Claude’s account does not change Codex’s.

Account cards show plan usage — Refresh re-checks sign-in and usage — and offer Reconnect when sign-in has lapsed. API-billed accounts — a Claude Console login or an OpenAI key — show what they’ve spent today and this month instead of plan meters; the Codex figure is estimated from tokens. The card’s menu renames or removes an account; you can’t remove the one that’s active while another exists, and a card whose turns are still running waits for them.

Models and effort

Inside an agent you choose a model, and for most models an effort level — how much thinking the model does before it answers. Raven carries its own catalog for each agent and adds models as they ship, tagged New for their first few days. A fresh install enables a short current set; the picker shows the ones you’ve enabled, and + Add models… at its foot opens Settings › Chats › Chat models › Enabled models for the rest. Codex models your plan can’t run are grayed out with a Not on your plan chip.

Effort is set per model, from its row in the picker: the level chip on the right opens a menu. Current Claude models run low, medium, high, xhigh, max (older ones have fewer levels, or none); Codex adds ultra on its top models, and the same menu sets a Codex model’s speed tier (Standard or Fast) and context window. Some models have no effort setting and the picker doesn’t show one.

The defaults are deliberate. New Claude chats start on Opus 5.5 at medium; when Codex is the only agent signed in, they start on GPT-6.1 Sol at high (or your plan’s own default model, if the plan can’t run Sol). Everything else starts at high — genuinely useful without being slow — and the top tiers stay something you reach for on purpose. Set your own default model and mode in Settings › Chats › New chat defaults.

Modes

Mode is the dial that decides what the agent is allowed to do. It sits in the composer, ⇧Tab cycles it, and you can change it mid-conversation.

Mode	What it does
Ask	Conversational. Reads the codebase, searches, looks things up in connected tools, and looks at the browser and simulator you have open; makes no edits.
Plan	Proposes a plan without editing application code. Can save plan documents and create or update work in connected Linear; looks at, but doesn’t act in, the browser and simulator.
Build	Executes — writes files, runs commands, writes design boards and context docs, drives the browser and the iOS simulator, signs in with your 1Password logins, sets up the Run button.

The useful habit is Ask or Plan first on anything non-trivial, then Build once the approach looks right. Ask is for understanding a codebase; Plan is for agreeing on an approach before implementation; Build is for doing it. Sketches, cards, and the pictures the agent shows of its work all work in every mode. With Codex, Ask and Plan run in a read-only sandbox. Whether a Build-mode agent reaches for the browser on its own is Settings › Chats › New chat defaults › Agent browser use.

Mode is per chat tab, so one tab in a workspace can be exploring in Ask while another is building. New chats start in Build unless you change the default.

Usage meters

The composer can show two gauges, for both Claude and Codex:

• A context ring — how full the current conversation’s context window is. It turns yellow at half and red near full; when it fills, start a fresh chat or clear the one you’re in rather than fighting it.
• A plan usage meter — how much of your provider plan’s allowance the current window has used, following the 5-hour and weekly limits plus the limit for the model your chat is on. Codex shows workspace credits here too. Turn it off in Settings › Chats › Chat input.

Both read from your provider account, and API-billed accounts show spend instead. Raven doesn’t bill for model usage — you’re using your own Claude or Codex plan.