# AI coding agent in Slack > An AI coding agent the whole team steers in Slack: the real Claude Code in a cloud sandbox per change, live previews anyone comments on, and a merge an engineer presses. Box hours by plan, your team's API key or your own Claude subscription. Source: https://heyray.io/coding Ask in a Slack thread and the real Claude Code builds the change. Anyone on the team comments on its live preview, and an engineer merges it. Each person can sign in with their own Claude subscription, or run on the team's Anthropic API key. ## How a change ships 1. **Say it in Slack.** Becca asks Ray in the thread that holds the brief. Ray opens a Slack code channel for the change and brings the brief with it. 2. **Everyone steers one agent.** Each message in the channel reaches the agent with the sender's name on it. Its questions and plans come back as buttons anyone there can answer, and anyone can stop it. 3. **Point at what's wrong.** Each service gets a live preview. Megan clicks the headline, writes what should change, and the fix reloads in the same preview. 4. **An engineer merges it.** Ray opens one pull request that names who asked and who steered. Only a developer or admin can merge it, and Dan does it from the card in Slack. ## It knows your team - **Ray's memory.** Searches what Ray remembers about your team, so you don't explain the background twice. - **Your tools, read as you.** Reads Linear, Notion, PostHog and your other connected tools as the person whose turn it is. - **Changes with an Undo.** A change it can take back lands with an Undo. A send waits for a yes. - **It asks the team.** Posts to the team, and asks the channel when it needs a decision. ## Who it's for - **Technical teams.** The real Claude Code, in a sandbox built from your repository. Steer it from Slack, or open Ray Workshop for the diff and your own terminal. Sign in with your own Claude subscription, and your turns run on it. Merge from a card in Slack. - **Marketing, ops and agencies.** Ask for the change in Slack, in your own words. Click the live page and say what should change. An engineer merges it. You never need a GitHub account. A client can see it first on a view-only link that ends in 7 days. Every full member of your Slack can start and steer a change and comment on the preview. Developers and admins also get the terminal and merge. Guests and Slack Connect partners never start or steer. ## Limits the model can't change - The sandbox never holds a GitHub token. Its git goes through Ray. - The sandbox never holds your Anthropic API key. Ray adds it to each call outside the sandbox. - Only staging and test values go in, never production credentials or databases. - Each teammate works in the sandbox as their own user. - The network is an allowlist. Any other host waits for a developer or admin to press Allow. - Nothing merges until a developer or admin presses Merge PR. ## Pricing Ray's price stays flat by team size, with no seats. On top of it, coding counts box hours while a change's sandbox is awake, and its model use runs on your team's API key or your own Claude subscription. - **Free.** $0, with 5 box hours, once, and they never expire. - **Ray Pro for 1 to 3 people.** $49 a month, with 40 box hours a month. - **Ray Pro for 4 to 10 people.** $99 a month, with 120 box hours a month. - **Ray Pro for 11 to 25 people.** $199 a month, with 300 box hours a month. **Box hours.** A sandbox sleeps after 10 idle minutes and stops counting. At the cap, new changes wait for next month, and changes already running keep going. **Your team's API key.** An admin adds your workspace’s Anthropic API key, and Anthropic bills you for what it uses. Ray caps it at $10 a change and $100 a day by default. An admin can change those caps. When a change hits one, a developer or admin can press Raise to keep it going, and a turn already running always finishes. **Your own Claude subscription.** Anyone can sign in to Claude Code once, from Ray Workshop, with their own Claude subscription. Their turns run on it from then on, and Workshop shows their 5-hour and weekly limits beside the key’s spend. A teammate who hasn’t signed in types as usual, and their turns run on the team’s key. ## What teams ask first **Which coding agent does it run?** The real Claude Code, unmodified, on Opus 5.5 by default. Ray builds the cloud sandbox around it and runs the channel, the previews and the merge. **Can people who don't code ask for changes?** Yes. Any full member of your Slack asks in a thread in their own words, then clicks the part of the live preview that should change and says what to fix. The real Claude Code does the work in a sandbox, so they never open a terminal or need a GitHub account. A developer or admin merges the pull request. **Can it merge without us?** No. Ray merges only after a developer or admin with write access presses Merge PR on its card in Slack. Ray enforces that in code, and the agent can't change it. **Does it use Slack's code channels?** Yes. Ray opens a Slack code channel for each change and brings the asking thread's context with it. Anyone in the channel can steer the agent or stop it, and Slack posts a card for the channel in the thread you asked in. **Does the agent keep working when nobody is watching?** Yes. Each change runs in its own cloud sandbox, so the agent keeps working with every laptop closed and posts to the change's channel when it needs a decision. The sandbox sleeps after 10 idle minutes and wakes on the next message or preview visit. **Can Ray answer a question about our code without a change?** Yes. Ask in a thread and Ray reads your repository to answer, with links pinned to a commit. If the answer needs real work, it offers a Start this change button. **What does it cost?** Coding comes with every Ray plan: 5 free box hours to start, then 40, 120 or 300 a month on Ray Pro, by team size. Model use runs on your team's Anthropic API key, capped at $10 a change and $100 a day by default, or on a person's own Claude subscription. **How do we keep spend under control?** Ray caps the team's API key at $10 a change and $100 a day by default, and an admin can change both. When a change reaches a cap, a developer or admin can press Raise to keep it going, and a turn already running always finishes. Box hours count only while a sandbox is awake. **Can we use our own API key or Claude subscriptions?** Both. An admin adds your workspace's Anthropic API key, and Anthropic bills you for what it uses. Anyone can also sign in to Claude Code once, from Ray Workshop, with their own Claude subscription, and their turns run on it from then on, with their 5-hour and weekly limits shown. A teammate who hasn't signed in runs on the team's key. **Where does the code run, and what can it reach?** In a cloud sandbox Ray builds for each change from your GitHub repository. It holds staging and test values only, never a GitHub token or your API key, and any host outside its allowlist waits for a developer or admin to press Allow. **What happens when we run out of box hours?** New changes wait for next month, and changes already running keep going. Hours count only while a sandbox is awake.