Coding agents

Each stage runs one agent: a harness, a model and a skill.

• Personal subscription: account connections through Agents → Your personal subscriptions in local and hosted Bento. Applies to matching models on stage runs you start; see setup and limits. Other local login-sharing options are described in the per-tool notes.
• Steer: what happens to a message you send while the agent works. Yes means it reaches the turn already running. Queued means it waits in the same session for the current turn to end. "Next run" means it waits for the run to end and starts a new one, resuming the conversation where the tool can.
• Interrupt: Stop ends only the current turn, and the agent waits for your next message in the same session. Where it is No, Stop ends the run.
• Questions: the agent can ask you a question on the card and wait for your answer.
• Cost: whether the tool reports a price for the run. "Tokens only" means it reports token counts and no price.
• Interrupt and questions apply to a card's stage runs (see Talking to a working agent). A Codex profile with codex exec flags, or an opencode profile with extra arguments, runs the one-shot command: no steering, interrupt or questions, messages wait for the next run, and opencode reports no tokens.
• Credentials for each tool are set under Agents in the web console.

Keys are stored encrypted and are personal by default, whether you add them in the Web UI or with bento setup. Existing shared keys keep their scope. See Personal and shared API keys. In multi mode, and in local mode with the Docker sandbox, a sandbox gets a proxy token in place of most keys, and the model proxy adds the key on Bento's side (see Security).

OpenRouter: pick an OpenRouter model on Codex, pi, or opencode and save OPENROUTER_API_KEY. Claude Code still needs the OpenRouter key saved as ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL set to https://openrouter.ai/api/v1.

DeepSeek: use pi or opencode with DEEPSEEK_API_KEY or openrouter/deepseek/... for streamed runs. DeepSeek Harness (dsh) is a preview.

Warm sandboxes reinstall pi below 0.70.1, opencode below 1.14.24, or dsh when its version does not match the pin.

Export and import agents as YAML from Agents, Settings, Config, or bento agents export / import. See the agents file.

Personal and shared API keys

Open Agents, Model provider keys, select the provider, paste the key, and save. New keys and endpoint settings belong to you. With beta features on, you can also save your own key for a custom provider under Agents, Custom provider keys after an owner or admin defines the provider.

Owners and admins can turn on Shared organization key before saving. With the toggle off, the field is labeled Personal API key. Turning the toggle on shares that key with the current team. The toggle chooses where the next save goes; it does not move an existing key. To change an existing key's scope, save it in the scope you want and remove the old copy. Members can always manage their own keys, but not shared keys or anyone else's personal keys. Only share keys licensed for team use; a subscription key meant for one person must stay personal.

Stage runs and stage follow ups you start use your personal keys before shared keys. A personal key uses only your own base URL for that provider, never the team's, so save a custom base URL in the same scope as its key. Teammates cannot steer a run that uses your personal key. Background stages, judges, conflict resolution, swarms, and remote runners keep using shared or machine keys; they never borrow the project owner's personal key. The GitHub connection stays shared.

Deleting a personal key stops later proxied requests that use it. Saving the key again does not revive an old run or switch it to a shared key; start a new run. Replacing an existing key's value rotates it for later proxied requests. Keys for providers that cannot use the model proxy still reach the agent directly, so removing one from Bento cannot take back a value already handed to a running agent.

Talking to a working agent

Messages

A message sent while a live agent works:

• Claude Code: queued, then runs as one more turn. Several messages become one turn.
• Codex: cancels the model call in flight. The next call includes the message.
• opencode: read at the next model call, in the same turn.
• pi: delivered after the current tool call.

A message to a live agent between turns starts a new turn. Other agents get the message when the run ends: Cursor, Antigravity, Muse Code and fx resume the session, while pool and dsh start fresh with the stage prompt and a compacted transcript. Bento does the same when a session is lost.

Stop

On Claude Code, Codex, opencode and pi, Stop ends the current turn, like Escape in a terminal. The agent waits for your next message in the same session. Stop works from the moment a run starts and is disabled while the agent waits.

The transcript shows "turn stopped" (not a failure), then:

If no follow up message is sent after 90s the agent run will end. Sending a follow up message will restart the stage.

If you send nothing within the stage's idle window (90 seconds by default), the run closes as stopped, not failed, and the card stays in its stage. Cost, tokens, artifacts, pushed commits and the sandbox are kept. The transcript shows:

Session paused. Your next message will resume it.

Stop ends the run on judge, conflict resolution and swarm runs, and on agents without interrupt. Right after a server deploy, a Codex or pi turn already in progress cannot be interrupted: Stop ends that run but keeps the session.

Questions

Claude Code, Codex and pi can ask questions on the card (pi through an ask_user tool Bento installs). The agent waits for your answers.

Questions appear above the composer, one at a time. A single choice answer moves on by itself. Otherwise press Next or Enter. Back, or a question's chip, returns to an earlier answer. With several questions, a review step lists every answer before Send.

A question disappears once the agent stops waiting for it. Unattended runs (swarms, judges, conflict resolution) decline questions, so the agent decides and says what it chose.

Personal account connections

Connect your own Claude or ChatGPT subscription under Agents → Your personal subscriptions. Setup is the same in local and hosted Bento. On hosted Bento, sign in with your own Bento user account before connecting. Each person connects separately; companies and teams must not share subscription logins.

Connect Claude

1. Open Agents → Your personal subscriptions and click Connect Claude.
2. Open the Claude sign in link, sign in to your own Claude account, and authorize the connection.
3. Copy the complete authorization code, including the # and the text after it (code#state).
4. Paste it into Paste the complete authorization code in Bento and click Finish connection.
5. Wait for Claude to show Connected to your account.

This one connection supports both Claude Code and pi with Anthropic models.

Connect ChatGPT

1. Open Agents → Your personal subscriptions and click Connect ChatGPT.
2. Open the ChatGPT sign in link and sign in to your own ChatGPT account.
3. Enter the device code shown in Bento and approve the connection. If the provider refuses device-code login, enable it in your account's security settings and start again.
4. Keep Bento open while it checks for approval. When complete, ChatGPT shows Connected to your account; there is no authorization code to paste back into Bento.

This one connection supports both Codex and pi with models available through Codex.

Start a run

Choose a matching agent and model, then explicitly start a card's stage run or send a follow-up message. A connected subscription takes priority for matching models, so you do not need a separate provider API key for those runs. In Codex, a provider-prefixed model such as openai/gpt-5-mini selects OpenRouter, not your ChatGPT subscription.

Custom provider endpoints keep using their API keys. Background stages, judges, conflict resolution, swarms, and remote runners keep their existing credentials. They do not borrow your subscription or the project owner's login.

Your connection belongs to you, even inside a team. Other members, including owners and admins, cannot list, change, or use it. They cannot steer, answer questions on, or change your card while your personal subscription run is active. Follow-ups from different people are handled in separate runs.

Reconnect or disconnect

Return to Agents → Your personal subscriptions and click Reconnect if Bento asks you to sign in again. For an expired device code, start a new connection. For an invalid Claude authorization code, paste the complete code#state from the current login attempt. Cancel connection cancels a pending login.

Click Disconnect to stop using a connection. Its next model request is refused; start a new run to use your configured API credentials. A running subscription task never silently switches to paid API credentials. Reconnecting creates a new connection, so start a new run after reconnecting rather than trying to revive an old task. Automatic retries retain the original user and connection.

Bento encrypts your provider tokens and refreshes them on the server. Agents receive only a temporary Bento run token through the model proxy. Provider subscription limits and Bento hosting charges still apply; a reported subscription model cost is a notional list price, not an API charge.

Claude Code on a subscription

Use Personal account connections for browser login in local or hosted Bento. The older setup-token option below is local mode only. Hosted organizations cannot save a shared Claude subscription token.

claude setup-token

Save the token under Agents, Claude subscription, in bento setup, or as CLAUDE_CODE_OAUTH_TOKEN= in .env (docker compose). The Web UI value overrides .env.

With a token saved, ANTHROPIC_API_KEY is not sent. ANTHROPIC_BASE_URL forces API key use, since tokens only work at Anthropic's endpoint.

Do not use macOS Keychain login on servers: its tokens rotate and are unavailable in containers. On "OAuth access token has been revoked", run claude setup-token again and save the new token.

Per tool notes

Claude Code

Bare model ids (claude-sonnet-5, claude-opus-5).

pi

provider/id model ids. Use a personal Claude or ChatGPT connection for eligible runs, or the key for the model's provider (ANTHROPIC_API_KEY, OPENAI_API_KEY, OPENROUTER_API_KEY, GEMINI_API_KEY or DEEPSEEK_API_KEY).

Codex CLI

Bare ids (gpt-5-codex) or provider prefixed ids (openai/gpt-5-mini, vercel/moonshotai/kimi-k3). Bare IDs can use your personal ChatGPT connection for eligible runs, or OPENAI_API_KEY. Provider-prefixed OpenRouter and Vercel models use their respective API keys.

Cursor CLI

Model ids per your Cursor plan; unlisted ids can be typed in. CURSOR_API_KEY. Runs headless.

opencode

provider/id model ids, including openrouter/. Same keys as pi.

Poolside (pool)

poolside/laguna-s-2.1 and other Laguna ids. POOLSIDE_API_KEY. Bento sets the model through POOLSIDE_STANDALONE_MODEL; override the endpoint locally with POOLSIDE_STANDALONE_BASE_URL. Alternative: pi or opencode with openrouter/poolside/laguna-s-2.1.

DeepSeek Harness (dsh)

Preview, pinned to @deepseek-ai/dsh@0.1.1-rc.2. Bare ids (deepseek-v4-pro). DEEPSEEK_API_KEY, optional DEEPSEEK_BASE_URL. Prints only the final message, so use Changes to see file edits.

Antigravity CLI

Google's agy, run headless. Model slugs name the tier and reasoning effort (gemini-3.1-pro-high, gemini-3.6-flash-medium); unlisted slugs can be typed in.

Uses GEMINI_API_KEY, optionally with GOOGLE_GEMINI_BASE_URL, so only Gemini models are available. The Claude and GPT models need a signed in Google account, which local mode can share from this machine (Agents, "Use this machine's logins"). Sharing a login carries the usual risk.

Antigravity bills against a plan's quota, not per run. MCP servers are written to ~/.gemini/config/mcp_config.json before every run.