NewConnect Sentry, PostHog and your database

A coding agent that runs on the model you already pay for

Sinatra takes work from Linear, GitHub, Slack, Telegram, the CLI or an MCP client, writes the code in its own cloud sandbox, runs your tests and opens a pull request. The model behind the agent is yours: a Claude, ChatGPT, SuperGrok or Cursor subscription you connect, or an Anthropic or OpenRouter API key your workspace shares. We never bill for model usage.

Sinatra runs the agent, you supply the model

Sinatra is the orchestration layer: the sandbox, the checkout, the conversation with you and the pull request. The model is the part you choose, along with how it's paid for. Every account needs its own model credentials, and that includes paid Sinatra plans. Your provider's charges and usage limits apply separately from your Sinatra plan.

You can connect more than one credential. What a task runs on depends on your own default, the workspace default and what you and the workspace have connected, in that order, and you can see the result for any repository before you start.

What you can connect

Claude subscription

Each member connects their own, in setup or Settings. It runs that member's tasks on Claude models.

ChatGPT / Codex subscription

Each member connects their own, in setup or Settings. It runs that member's tasks on the supported OpenAI models, the openai-codex/* entries.

SuperGrok subscription

Each member connects their own, in setup or Settings. It runs that member's tasks on xai-oauth/* models, on the oh-my-pi engine.

Cursor subscription

Each member connects their own, in setup or Settings. It runs that member's tasks on cursor/* models, on the oh-my-pi engine.

Anthropic API key

An owner or admin pastes it for the workspace. It runs authorized members' tasks on Claude models.

OpenRouter API key

An owner or admin pastes it for the workspace. It runs authorized members' tasks on the openrouter/* models Sinatra lists.

Connecting ChatGPT uses OpenAI's device-code sign-in: you copy a code, enter it on OpenAI and approve. OpenAI asks for a paid ChatGPT plan (Plus, Pro, Business, Enterprise or Edu) with Device code authorization for Codex turned on in ChatGPT Settings → Security. On a Business, Enterprise or Edu workspace an admin turns it on. You can close the window after approving, because Sinatra keeps checking for the 15 minutes the code lasts.

There is no OpenAI API-key field. An OpenRouter key goes with the openrouter/* models Sinatra lists, and it doesn't stand in for a personal Codex, SuperGrok or Cursor connection. The full list of models, with the reasoning effort levels each one supports, is in the models docs.

Personal subscriptions and shared workspace keys

A personal subscription belongs to one member in one workspace and runs the tasks that member starts. Connecting yours gives nobody else access to it, so on a team each person who wants to use a subscription connects their own.

API keys are the shared option. An owner or admin pastes an Anthropic or OpenRouter key in Settings, and the workspace's members can run on it. When your personal subscription can serve the selected model, it's used ahead of the shared key. Pasted keys are encrypted per workspace, and the encryption is bound to that workspace, so a key stored in one workspace can't be decrypted in another.

How the model for a task is picked

Three settings decide it, from highest priority to lowest. Your default, under Settings → Models → Your default model, is an engine, a compatible model and optionally a reasoning effort, saved by any member for the work they ask Sinatra to do in that workspace. It applies in every repository. The workspace default is the same choice made by an owner for everyone. Only an owner can change it, and Reset to automatic removes it. When neither is saved, Sinatra selects a model from the credentials available to you.

Automatic selection goes in this order. A personal Claude subscription runs anthropic/claude-opus-5-5 at medium effort. A personal Codex subscription runs the newest Sol model your ChatGPT account offers, ahead of any API key. A single Anthropic key runs anthropic/claude-opus-5-5, and a single OpenRouter key runs openrouter/anthropic/claude-opus-5.5. With several keys connected, Sinatra picks the alphabetically first provider that has a usable default. SuperGrok and Cursor come last. With none of those, SuperGrok runs xai-oauth/grok-4.7, and with no SuperGrok either, Cursor runs cursor/grok-4.7. To run on one of them while you have another credential, save it as your default.

Reasoning effort travels with the default that chose the model. A default saved without an effort runs at medium. To check before you start, open a repository and choose the Agent tab: Next task here shows the engine, model and effort your next task there will use and which setting chose each. A task started from the CLI, MCP or API can also name its own model, which takes priority over every setting above for that task.

A change applies to agents started afterwards. An agent already running keeps the model it started with. Changing engines starts a fresh agent conversation, because one engine can't resume another's history, but the task itself stays.

Settings choose the model, the repository sets up the work

Repositories no longer choose the engine, model or effort. Sinatra rejects a .sinatrarc.json that contains engine, agent.model or agent.effort, whatever the value, and the error names the keys to remove. The model is a person's choice in Settings, so it doesn't change when a task's repository does.

The file, committed at the root of the repository, still sets up the work there. hooks.test_command runs after the agent finishes, and Sinatra re-runs the agent to fix failures up to twice. If tests still fail, it reports the failure, and the pull request isn't blocked. sandbox.env holds plaintext environment variables, sandbox.setup_commands.pre_clone runs when the sandbox is created if the repository is known by then, and ci.max_rework sets how many CI fixes Sinatra makes in a row before it stops and says so. The schema is in the .sinatrarc.json docs.

Secrets and standing instructions

Credentials your setup, tests or agent need go in secrets, never in the committed file. An owner or admin adds workspace secrets under Settings → Secrets, shared by every new task sandbox, or repository secrets under the repository's Access & secrets. Each secret needs the hosts allowed to receive it. In the sandbox the environment variable holds a placeholder, and the proxy swaps in the real value only on a request to an approved host. Model keys and subscriptions don't go here; they have their own connections in Settings → Models. Details are in the secrets docs.

Workspace agent instructions are a Markdown block, up to 16 KiB, that every agent in the workspace reads: cluster endpoints, dashboards, the commands your team uses. An owner or admin writes them under Settings → Agent instructions, and members can read them. They apply the next time Sinatra starts an agent. A repository's AGENTS.md and CLAUDE.md take precedence for that repository's conventions. See the agent instructions docs.

What it costs

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. Your Sinatra plan sets how many agents work at once. 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. The token and cost numbers Sinatra shows are estimates, not amounts owed to Sinatra, including when your subscription covers the usage. Every plan is on the pricing page.

Limits and caveats

Your provider's usage limits still apply, and a subscription you connect runs only your own tasks. ChatGPT plans don't all include the same Codex models, and lower plans often lack the newest one, so pick a model your plan lists. On Cursor's X Subscription plans, Claude and GPT models through Cursor are billed at API prices through top-ups, while Grok models use the plan's included usage. That's why Cursor's automatic choice is a Grok model. Choose cursor/claude-opus-5-5 or cursor/gpt-5.6-sol yourself if you accept that cost.

SuperGrok and Cursor run only on the oh-my-pi engine. Sinatra accepts the engine and model pairs Settings offers, which are not every model each provider sells. Saving a default doesn't provide the credential it needs, and a missing credential for the selected model blocks the run.

Common questions

Can my teammates use my Claude or ChatGPT subscription?

No. A personal subscription belongs to one member in one workspace and runs only the tasks that member starts. To give the whole team model access, an owner or admin adds an Anthropic or OpenRouter API key, which is shared with the workspace's members.

Can I paste an OpenAI API key?

No. Sinatra has no OpenAI API-key field. OpenAI models run through a personal ChatGPT / Codex connection. An OpenRouter key reaches the openrouter/* models Sinatra lists.

Can I set the model in .sinatrarc.json?

No. Repositories no longer choose the engine, model or reasoning effort, and Sinatra rejects a .sinatrarc.json that contains engine, agent.model or agent.effort, whatever the value. Remove those keys and pick your model under Settings → Models → Your default. An owner can also set a workspace default.

Does a paid Sinatra plan include model access?

No. Every plan, paid ones included, needs your own model credentials. 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. Your provider's charges and usage limits apply separately from your Sinatra plan.

What happens if I haven't connected anything?

Your tasks are blocked with own_credentials_required until you connect a personal subscription or an owner or admin adds a workspace API key. Choosing a model in Settings doesn't create the credential it needs either: a missing credential for the selected model blocks the run.

Once a model is connected, tasks start from a Linear assignment, the Sinatra label on a GitHub issue, a Slack mention, an MCP client or the CLI and API.

Start free and connect a modelPro is $20 a month for the whole workspace