A coding agent that works in its own cloud sandbox
Hand Sinatra a task from Linear, GitHub, Slack, Telegram, the CLI or an MCP client. It checks out your repository in a sandbox we host, sets the project up, makes the change, runs your tests and opens a pull request. Your laptop isn't involved, and you can hand it the next task while this one runs.
What runs where
Each task gets its own sandbox: a checkout of the repository and a place for the agent to run commands and tests. The agent works there, commits to a branch and pushes it to GitHub. Its progress and replies go back to where the task started, the ticket, the Slack thread or the Telegram topic, and a task from the CLI, MCP or the API keeps its conversation in Sinatra.
The sandbox belongs to the task. When you reply, the next agent picks up the same sandbox and branch where it can, so a follow-up doesn't start from a fresh checkout. A question that needs no code is answered without checking anything out.
From ticket to pull request
You assign a Linear issue, add the Sinatra label to a GitHub issue, mention @Sinatra in Slack, message the Telegram bot, or start one from the CLI, MCP or API. The agent reads the whole issue or thread first. If the request is vague it asks questions or proposes a plan and waits for your reply.
If the repository is known when the sandbox is created, its sandbox.env values, repository secrets and pre_clone commands from .sinatrarc.json apply then. New sandboxes start from a generic base image unless your file names a snapshot.
The agent clones the repository itself, or reuses a checkout it has already verified. It reads .sinatrarc.json, your README, AGENTS.md, manifests and any snapshot Dockerfiles as setup hints, then installs dependencies and any tools the base image lacks.
The agent makes the change, verifies it, commits, and opens a draft pull request.
If you set hooks.test_command, Sinatra runs it once the agent finishes. When tests fail, the agent gets up to two tries to fix them. If they still fail, the failure is reported and the work continues.
Sinatra reviews its own diff against the ticket, tries to mark the draft ready, and links it on the ticket. The sinatra / review check reports what the review found when the GitHub App can write checks.
The task doesn't end there. Request changes or comment on the pull request and the agent pushes revisions to the same branch. A failed GitHub check suite on its pull request starts an agent to fix it, up to ci.max_rework fixes in a row (two by default) with no message from a person.
What you configure
Nothing is required. When you want more control, commit a .sinatrarc.json to the repository's default branch. Sinatra reads it from there.
hooks.test_command is the command Sinatra runs after the agent finishes, with the two tries at fixing failures. If you need more than one check, combine them in this one command.
sandbox.setup_commands.pre_clone runs when the sandbox is created. It runs once per sandbox, not before every follow-up, and changing it doesn't reinstall anything in a sandbox the task already has. Steps that need the checkout belong in your README or AGENTS.md, where the agent reads them.
sandbox.env holds plaintext, non-secret environment variables. It is committed with the file and the agent can read it. sandbox.snapshot_name selects an existing snapshot instead of the base image; for a custom one, contact us about registering it first.
API tokens go in repository or workspace secrets, set by an owner or admin, each with the hosts allowed to receive it. Every sandbox command also sees SINATRA_SANDBOX=1, so a setup script, git hook or test config can tell it is running in Sinatra. The engine and model don't live in this file; you pick them in Settings.
Isolation and credentials
Sandboxes are never shared between workspaces. Before reusing one, Sinatra checks it belongs to your workspace, and a sandbox that can't be proven yours is treated as not found, so the task creates its own.
All sandbox traffic goes through our egress proxy. Public hosts are reachable and audited; private, loopback and cloud metadata addresses are blocked. Secrets reach the sandbox as placeholders. The proxy swaps in the real value only in a request header to one of that secret's approved hosts and rejects it anywhere else. Model keys work the same way.
Sinatra works on branches and opens pull requests. Its workflow doesn't merge or deploy, so branch protection and required reviews still apply. Your scripts and the credentials you give it decide which of your services its commands can act on, so scope them to what the task needs. The security page has the details on encryption and data handling.
Limits and caveats
Failing tests don't block the pull request. After two tries the failure is reported and the pull request can still be marked ready, and a self-review error can still end in a ready pull request too. Neither guarantees green CI.
Each agent works up to 60 minutes on Free. At 60 minutes the agent stops and saves its work. Upgrade to let it finish. The work is saved to the task's branch and a draft pull request where it can be. Agents work until they're done on paid plans. On every plan, an agent that makes no progress for 30 minutes is stopped with its changes kept, and a reply carries on.
Between replies the sandbox, the branch and the agent's session are kept where they can be. With the oh-my-pi engine, a reply within five minutes continues with the agent still warm, and a later reply resumes the saved session after a cold start. A sandbox is stopped when work finishes, stops after two hours idle, is archived an hour after that, and is deleted after 30 days without activity. Commit anything that has to survive: uncommitted files and the agent's conversation aren't guaranteed to outlive a lost sandbox.
Secrets are placeholders meant for HTTP header authentication, so a tool that needs the raw value locally won't work with one. Repository secrets and sandbox.env are applied when the sandbox is created and the repository is known. If the agent picks the repository later in the conversation, start a new task that names the repository when its setup or tests need those secrets. The sandbox disk is finite too: a task that fills it reports the full disk, and the same workload can hit it again.
What it costs
The sandbox comes with every plan. Free is 1 agent at a time, plus 1 waiting, with no daily limit and no card. Pro is $20 a month for the whole workspace: 3 agents at a time, plus 10 waiting. Bring your own Claude, ChatGPT or Cursor subscription, or an API key. Sinatra never bills for model usage. That's between you and your provider.
Common questions
Does anything run on my machine?
No. Every task, including one started from the sinatra CLI or an MCP client, runs in a sandbox Sinatra hosts. The agent checks out your repository there, runs commands and tests there, and pushes a branch to GitHub. What you get back is a reply on the ticket or thread and, when the work changes code, a pull request.
Do I have to configure anything before the first task?
No. The agent clones the repository itself and reads your README, AGENTS.md, CLAUDE.md, manifests and .sinatrarc.json to work out how to install and test the project, and it can install the tools it needs. A committed .sinatrarc.json is optional. Add one when you want Sinatra to run a specific test command after the agent finishes, or to run commands when the sandbox is created.
What happens when my tests fail?
If .sinatrarc.json sets hooks.test_command, Sinatra runs it after the agent finishes, usually with a draft pull request already open, and gives the agent up to two tries to fix failures. If the tests still fail, Sinatra reports the failure and carries on. A failing test command doesn't stop the pull request from being marked ready, so read the test and review results before you merge. Your repository's required checks and your own review stay the merge gate.
Can the sandbox reach the internet, or my private services?
Sandbox traffic goes through our egress proxy. Public hosts are reachable and audited. Private, loopback and cloud metadata addresses are blocked at the proxy, so the sandbox can't reach a service on a private network. Public egress means repository content can leave the sandbox in a network request; we detect unusual requests, we don't prevent them.
How do API keys and other secrets get into the sandbox?
Add them as repository or workspace secrets in the dashboard, each with the hosts allowed to receive it. They're stored encrypted. In the sandbox the environment variable holds a placeholder, and the proxy swaps in the real value only on an HTTP request header to an approved host. A tool that needs the raw value locally, for example to sign data or to send it in a request body, can't use a placeholder.
What happens to the sandbox when the task is done?
Sinatra stops it when the work finishes. An idle sandbox stops after two hours and is archived an hour after it stops. A reply to the task can reuse it. An archived sandbox that goes 30 days without activity is deleted by a daily sweep. A sandbox is not a backup: anything the agent didn't commit to the branch can be lost with it.
How long can one agent work?
Each agent works up to 60 minutes on Free. At 60 minutes the agent stops and saves its work. Upgrade to let it finish. Agents work until they're done on paid plans. On every plan, an agent that makes no progress for 30 minutes is stopped with its changes kept, and a reply carries on.
The sandbox docs cover setup, retention and network access in full. Tasks start from a Linear assignment, the Sinatra label on a GitHub issue, a Slack mention, an MCP client or the CLI and API.