Sinatra vs Claude Code
Anthropic's agent, compared here on the two surfaces that open pull requests while you are elsewhere: the Claude Code GitHub Action and Claude Code on the web. Not the terminal CLI.
Sinatra runs Claude Code. That is not a rhetorical concession: a Sinatra task on a Claude credential is Claude Code doing the work, in a sandbox we operate, driven by a tracker instead of by you. So this page is not about which agent is smarter. It is about the wrapper. Claude Code on GitHub means a workflow file and an auth secret in every repository, no Linear delegation, and no tier that runs on Anthropic's key for you. If you are one developer on one repository, Claude Code alone is cheaper and you should use it.
- Yes means the vendor's own app starts a run from the issue itself, with nothing to commit to the repo first. Sinatra's trigger is the Sinatra label that ships with the GitHub App install (`sinatraai` on the hosted app). Partial means either that a workflow file and an API-key secret have to be added per repository first, or that the mention only works on pull requests, or that starting from an issue needs a trigger or automation you configure yourself.
- Yes means assigning the ticket to the agent in Linear starts the run. Partial means the run starts some other way: an attachment link you click, or a trigger you configure yourself against Linear's webhooks. Assignment on its own does nothing.
- Yes means a documented HTTP API or MCP server starts a run on the vendor's hosted agent from a task the caller writes, so a script, a CI job, or an MCP client can hand off work and follow it to a pull request. Partial means only the vendor's CLI can start such a run, or the API only fires an automation saved ahead of time, or it is enabled for selected organizations only. Local agent sessions, and MCP servers that only read docs or tickets, don't count. Creating, labeling, or assigning an issue through a tracker's API, so that a tracker trigger or a saved automation picks it up, doesn't count either. Sinatra's MCP server takes an OAuth sign-in from Claude Code, Cursor, claude.ai or ChatGPT, or a personal access token; the HTTP API and the `sinatra` CLI take the token; and every run happens on Sinatra's hosted agent.
- Yes means a changes-requested review is enough on its own: Sinatra reworks the branch and pushes the follow-up commit with no re-trigger and no mention. Partial means the agent acts only when it is @-mentioned in the thread, or only while its original session is still open.
- Yes means your own provider API key runs the agent everywhere the product runs it. Partial means keys work on some surfaces but not the one being compared here, either because the vendor documents them as local-CLI-and-desktop only, or because it never documents them covering its cloud agent at all.
- Yes means a Claude Pro/Max or ChatGPT subscription you already pay for authenticates the run on the vendor's hosted cloud. Partial means it works only where that CLI's login already sits on the machine: your own laptop or a self-hosted backend, not a clean cloud sandbox, or that a linked subscription pays only for one provider's share of a hosted run while the vendor bills the rest.
- Yes means the vendor never charges you for model use itself: it runs on a key or subscription you pay a model provider for, or model use is included in a flat plan and capped by a rate limit or task quota rather than charged. When the vendor is itself the model provider, paying its own API rates on your own key counts as paying your provider. Partial means the vendor bills model use unless you bring your own key on the surface compared here. No means the vendor bills model use, as tokens, credits, or units, at cost or marked up, including usage past an included allowance. Sinatra runs on your own Claude or ChatGPT subscription or API key and never bills for model usage.
- Both halves have to be published: the plan price, and the rate for work past the plan. A published seat price sitting on top of an undisclosed credit, compute unit, or quota size is partial.
- Yes means the free tier states its limits as numbers. Sinatra's Free plan is 1 task at a time, plus 1 queued with no daily limit and no card, on your own subscription or API key; each request runs for up to 60 minutes, and if it needs longer Sinatra stops, keeps the changes, and picks up when you reply. Partial means a free plan exists but its limits are not published, or are published in a unit whose size isn't. No means there is no free tier, or the free tier leaves out the agent compared here.
- Yes means the vendor supplies model access and bills for or includes that usage. Sinatra requires your own workspace API key or personal subscription on every tier, including existing and paid accounts, and does not bill model tokens.
- Yes means an editor plugin or a desktop app; partial means a terminal CLI and no plugin or app of the vendor's own inside an editor. Sinatra is partial: the `sinatra` CLI, and an MCP server that Claude Code, Cursor, claude.ai, ChatGPT and other MCP clients can call, hand work to the hosted agent, which works in its own sandbox and comes back as a pull request. There is no editor extension, and the agent never runs on your machine. If you want an assistant working inside your editor, that is a real reason to pick one of these instead, and several teams run both.
- Yes means mentioning the agent in a Slack channel or thread starts a run on the vendor's hosted agent. Sinatra's Slack app is connected once per workspace by an owner or admin, on every plan: mention it in a channel and it works in that thread, posting its progress and answer there and opening a pull request when the request needs code. Follow-ups in the thread continue the task.
- Yes means you can run the agent on infrastructure you control, or the source is open. Partial means execution or inference can be pointed at your own account or your own machines while the product itself stays the vendor's and stays closed. Sinatra is hosted only; there is no self-managed build.
Competitor facts were read off each vendor's own documentation and pricing pages on September 4, 2026. The “Start a hosted run over API or MCP” row was read on October 13, 2026. Sinatra's terms reflect the current own-credential requirement.
How you start it
- Add the Sinatra label to a GitHub issue. The label ships with the GitHub App install and is named after the app (
sinatraaion the hosted one) - Assign a Linear ticket to Sinatra, or @-mention it in a comment
- Mention @Sinatra in a Slack channel; it works in that thread, and follow-ups in the thread continue the task
- Start a task from your terminal with the
sinatraCLI, from Claude Code, Cursor, claude.ai or ChatGPT over MCP, or over the HTTP API. It runs on Sinatra's hosted agent, not on your machine
- @claude in a GitHub issue or pull-request comment, once you commit a workflow file and an auth secret
- Claude Code on the web, plus auto-fix on pull requests
- @Claude in Slack
- The Claude Code CLI, VS Code, and JetBrains
claude --cloud "…"from the CLI, signed in with a claude.ai account, which starts a cloud session (read October 13, 2026)- A routine's API trigger: a
/fireendpoint that runs the routine's saved prompt, optionally with run-specific text (read October 13, 2026)
Free: 1 task at a time, plus 1 queued, no daily limit. Bring your own Claude or ChatGPT subscription, or an API key. On Free, each request runs for up to 60 minutes. If it needs longer, Sinatra stops, keeps the changes, and picks up when you reply. No card. Paid plans are one flat monthly price for the whole workspace, however many people are in it: Pro $20 (3 tasks at a time, plus 10 queued), Team $99 (10 tasks at a time, plus 25 queued), Scale $249 (25 tasks at a time, plus 100 queued). Enterprise is custom. Sinatra never bills for model usage.
Sinatra pricing →You supply the credential: an Anthropic API key at published token rates, or a Claude Pro, Max, Team, or Enterprise subscription through a long-lived OAuth token, or your own Bedrock, Vertex, or Foundry account. GitHub Actions runs also consume your Actions minutes. There is no tier that runs on Anthropic's key on your behalf.
Claude Code pricing page →How you start a task
The GitHub Action route works well and it is per-repository. Anthropic's own docs say you configure one repository at a time, so an organization means an org-level secret plus a workflow file in every repo or a reusable workflow to call.
The web route is better: Claude Code on the web runs sessions on Anthropic's
infrastructure, and with the App installed, auto-fix responds automatically to CI
failures and review comments. Sinatra matches that rather than beating it. Both
also start a run from a script. claude --cloud "…" takes a task you write and
starts a cloud session, but it is a CLI that needs a claude.ai login on the
machine running it. A routine's /fire endpoint takes a token, and it runs the
prompt you saved in the routine, optionally with run-specific text. Sinatra's HTTP
API and MCP server take the task itself in one call, with a personal access token
or, for the MCP server, an OAuth sign-in.
Sinatra's setup is one App install for the organization. Nothing is committed to your repository and no credential sits in your repository secrets.
Linear is the clean split. Anthropic's platforms page lists Linear among MCP servers and connectors, which lets Claude use Linear inside a session it already started. There is no assign-to-Claude in Linear.
What it costs, in dollars
The model half is identical, because it is the same credential, and neither of us adds anything to your token bill.
The differences are at the edges. Claude Code on GitHub Actions burns your Actions minutes, since the agent runs on GitHub-hosted runners. Both products require your own model credentials. Sinatra requires them on every plan, Free included. Free is 1 task at a time, plus 1 queued, with no daily limit and no card. Personal subscriptions cover their owner's tasks within that provider's limits.
Sinatra's own charge, $20 a month for the whole workspace on Pro, buys the parts that are not the agent: the sandbox, the tracker integrations, the isolation, and not having a workflow file to keep current in every repository. For one developer on two repositories that buys very little you could not do in an afternoon. The maths changes at about the point where an afternoon becomes twenty repositories and a secret rotation policy.
Where Sinatra wins
Linear is a delegation surface. Assign the ticket, close the tab. Claude Code's Linear connector is not the same thing: it lets Claude read and update tickets in a session someone already started, and does not put the agent in a product manager's hands. (Sinatra's own MCP server points the other way: it lets Claude Code hand a task to Sinatra.)
Nothing lands in your repository: no claude.yml to maintain, no key in each
repo's secrets, no drift between twenty repositories' workflow versions. Runs
happen in Sinatra's sandboxes rather than on your runners, so the agent does not
compete with your CI budget. The free tier requires your own key or subscription.
And a task is a conversation rather than a single shot: reply on the ticket or
the pull request and the same task picks it up, even after a merge, and a message
sent while a run is going reaches the running agent.
Where Claude Code wins
Let us be direct: Claude Code is the better product, and Sinatra runs it.
It is the agent. Everything good about a Sinatra run on a Claude credential is Anthropic's work, and new capabilities land there first.
It works on your machine, through the CLI, VS Code, and JetBrains, on your
checkout with your uncommitted changes. Sinatra has no local mode and cannot help
you at 2am on a branch you have not pushed. Its sinatra CLI starts a task from
your terminal, but the task runs in Sinatra's sandbox on a fresh clone, not on your
checkout.
For one person on two repositories it is cheaper: install the App, commit one
workflow, keep the twenty dollars. And it gives you far more rope, with skills,
plugins, hooks, custom agents, --max-turns, your own runners, and inference
through your own Bedrock, Vertex, or Foundry account, where Sinatra exposes a
.sinatrarc. Auto-fix on the web already handles CI failures
and review comments, so do not switch for that row either.
Looking for a Claude Code alternative?
Very few people want an alternative to Claude Code. What they want is Claude Code somewhere it does not currently go: attached to a Linear board, or installed once across an organization instead of a workflow file at a time.
That is the entire pitch. Keep the model, keep the agent, keep the subscription,
and connect it to Sinatra so it runs your hosted tasks. What changes is the
wrapper: two tracker triggers plus a CLI and an MCP server, one org-wide install,
sandboxes we operate instead of your Actions minutes, and a .sinatrarc per
repository.
If your objection to Claude Code is the model or the quality of the diffs, Sinatra will not help you. It is the same agent doing the same work.
Switching, or running both
Running both is the normal end state, not a compromise. Keep Claude Code in your
terminal, because that is where you work on the branch you are holding, and hand
Sinatra the tickets nobody has started. You can also do that from inside Claude
Code: add Sinatra's MCP server and sign in with your Sinatra account, ask Claude to start
a Sinatra task, and it runs in Sinatra's sandbox and comes back as a pull request
while you keep working. @claude fires the Action, the Sinatra
label fires Sinatra, and the two never see each other's work. If you already have
a CLAUDE_CODE_OAUTH_TOKEN, pointing Sinatra at the same subscription takes a
minute and adds nothing to your token bill.
Questions people ask about Claude Code and Sinatra
Is Sinatra just a wrapper around Claude Code?
It runs Claude Code among other engines and does not hide it. What it adds is the wrapper: a tracker trigger with no workflow file, a sandbox per task, credentials encrypted per workspace, tasks that keep taking replies after the pull request, and per-repository configuration in a committed .sinatrarc.
Can Claude Code be assigned a Linear ticket?
Not natively. Anthropic lists Linear under MCP servers and connectors, which Claude can call during a session, rather than as a delegation surface.
What do I need to set up Claude Code on GitHub?
The Claude GitHub App, an auth secret, and a workflow file. Anthropic's docs note you configure one repository at a time, so an org means an org-level secret plus a workflow file in each repo.
Which is cheaper?
Claude Code, for one person on a couple of repositories. Sinatra adds $20 a month for the whole workspace on Pro, for the sandboxes and the tracker integration; the model bill is identical, because it is the same credential.
Where these facts came from
Every mark on this page comes from Claude Code's own documentation and pricing pages, not a third-party roundup: read on September 4, 2026, except the “Start a hosted run over API or MCP” row, read on October 13, 2026, as were the ways to start a run tagged with that date. Sinatra's side of this page was last updated on October 15, 2026. If something here is out of date or wrong, tell us and we will fix the page.