Sinatra vs OpenAI Codex
OpenAI's cloud agent: the one that runs tasks in ChatGPT and opens pull requests, not the Codex CLI on your laptop. Compared here on Codex cloud.
Codex cloud is genuinely good at the thing Sinatra does: assign a Linear issue, get a pull request. The catch is that everything runs through one vendor relationship. Your throughput is a ChatGPT plan tier, your models are OpenAI's, and the API-key path that would give you published per-token pricing explicitly turns the cloud features off. Sinatra runs on that same Codex subscription, but also on Anthropic, OpenRouter, or a Claude plan, and it starts from a GitHub issue label, which Codex does not.
- 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 October 16, 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
- Assign a Linear issue to Codex, or mention @Codex in a comment
- @codex in a GitHub pull-request comment, or a GitLab merge-request comment (beta)
- Mention @ChatGPT in Slack, which can hand repository work to a Codex cloud task
- Codex cloud in ChatGPT on the web, desktop, and mobile, plus the CLI and IDE extension
codex cloud execin the Codex CLI, which submits a cloud task directly- The openai/codex-action GitHub Action, from a workflow file you commit
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 →Included in a ChatGPT plan, with Codex cloud from Plus up: Plus $20/mo, Pro at $100, $200, or $500 a month, Business $20 a user a month billed annually ($25 monthly), Enterprise and Edu contact-sales. Free ($0) and Go ($8/mo) get Codex in the desktop app only. Usage is a rate limit rather than a rate: OpenAI estimates 15 to 160 local messages per five hours on Plus with GPT-6.1 Sol, cloud tasks may use more, and Pro currently has no five-hour limit, though weekly limits may apply. Past the limit, Plus and Pro can buy credits. An API key runs the CLI, SDK, and IDE extension at OpenAI's published API prices but unlocks no cloud features.
OpenAI Codex pricing page →How you start a task
On Linear these two are a genuine tie: install the integration, assign the issue, and it reports back on the ticket.
GitHub is where they part. OpenAI's integration is built around pull requests, so
there is no issue trigger in it. You can build one with the openai/codex-action
Action, which means committing a workflow file and storing an OPENAI_API_KEY
secret in every repository that needs it. Sinatra's label ships with the App
install and starts the run from the issue itself.
For a backlog in GitHub Issues rather than Linear, that is the difference between a one-click install and a per-repository project.
From a script, OpenAI documents starting a cloud task through the CLI's
codex cloud exec rather than an API, where Sinatra's HTTP API and MCP
server take the task directly, with a personal access token or,
for the MCP server, an OAuth sign-in. ChatGPT can sign in too:
on a ChatGPT Business, Enterprise or Edu workspace, on the web, a chat can hand
Sinatra a task. OpenAI limits custom MCP servers on Pro to reading, so a Pro
chat can't start one.
What it costs, in dollars
Codex is bundled, not sold, and the cloud agent comes with ChatGPT Plus and above. Free and Go get Codex in the desktop app only. What the plan does not come with is a rate. OpenAI publishes an estimate instead, which depends on the model: on Plus, 15 to 160 local messages per five hours with GPT-6.1 Sol, 5 to 45 with GPT-6 Astra, and far more with the small GPT-6 Luna. It adds the caveats that "cloud tasks may use more of your allowance than local messages" and that "these estimates are not fixed message limits." Two tickets that look the same to you can consume very different amounts. Past the limit, Plus and Pro users can buy credits, which OpenAI meters per million tokens, or move up a tier.
There is a path with real per-token pricing, and OpenAI's own pricing page says what it costs you: "Codex in the CLI, SDK, or IDE extension" with "no cloud-based features (GitHub code review, Slack, etc.)." Published pricing or the cloud agent, not both.
Sinatra publishes both numbers because they are separate things: the plan, $20 a month for the whole workspace on Pro, and your provider's rate for models. If the credential you connect is a Codex subscription, the marginal cost of a task is whatever that already costs you.
Where Sinatra wins
No lock-in to one model vendor. A Codex subscription works, and so does an
Anthropic key, an OpenRouter key, or a Claude plan, pinned per repository in a
committed .sinatrarc. You can change your mind next quarter without changing
tools.
A GitHub issue label starts a run, with no workflow file, no repository secret, and no Action to maintain across twenty repos. Review is the instruction, so asking for changes gets the follow-up commit without a mention; Codex's documented autonomous behavior is reviewing new pull requests, not acting on a review of its own. The free tier's limits are numbers rather than a range: 1 task at a time, plus 1 queued, with no daily limit. And a Sinatra plan sets how many tasks run at once, not how many you get: assign more and they queue, and the tokens they use are billed by your provider at its rate.
Where Codex wins
The models come first. Codex cloud runs OpenAI's newest models the day they land, tuned by the people who made them. If GPT is the family you want, nobody serves it better.
One subscription also covers everything: ChatGPT, Codex in the browser, the CLI, the IDE extension, iOS, Slack, and GitHub review, on one plan and one login, where Sinatra deliberately gives you two invoices. The web app and iOS are real surfaces Sinatra does not have at all. Codex reviews your team's human-authored pull requests on open, with security review as a first-class mode, which is a different job from closing tickets and one Sinatra does not do. And Pro, from $100 a month, currently has no five-hour limit, though OpenAI says weekly limits may apply.
Looking for an OpenAI Codex alternative?
Two things bring people here. Model lock-in: a team standardized on Claude cannot point Codex cloud at it, and running two agents to run two model families is silly. And the ceiling: when Plus stops being enough, the choices are credits or a $100 plan, and both buy more GPT rather than a different model.
Sinatra separates the two purchases. You pay $20 a month for the whole workspace on Pro for the machinery and buy models from whoever you want, including OpenAI, since a Codex subscription is a supported credential. Switching does not mean abandoning what you already pay for. Our plan sets how many tasks run at once; how much model work they do is bounded by your provider's limits.
If you are here because Codex's GitHub PR review is the part you like, keep it. It reviews your team's pull requests, which the two do side by side.
Switching, or running both
The triggers do not collide: Codex answers @codex on pull requests and a Linear
assignment, Sinatra watches a label and an assignee of its own. The experiment
worth two weeks is the one where both use the same credential. Connect your Codex
subscription to Sinatra, hand each five real tickets, and read the ten pull
requests knowing the model was identical. What you are measuring then is the
harness and the trigger, which is the only part that differs.
Questions people ask about OpenAI Codex and Sinatra
Can Sinatra use my ChatGPT or Codex subscription?
Yes, on the hosted cloud, alongside Anthropic and OpenRouter keys. The free tier is 1 task at a time, plus 1 queued with no daily limit and no card. Every account brings its own model credentials.
Can Codex start from a GitHub issue?
Not from OpenAI's GitHub integration, which is scoped to pull requests. You can wire issue events up with the openai/codex-action Action, which means a workflow file and a secret in each repository.
What does Codex cloud cost?
It comes with a ChatGPT plan from Plus ($20 a month) up, and what you get is a rate limit rather than a rate: OpenAI estimates 15 to 160 local messages per five hours on Plus with GPT-6.1 Sol, and cloud tasks may use more. Pro, at $100, $200, or $500 a month, currently has no five-hour limit, though weekly limits may apply. Past the limit you can buy credits or move up a plan.
Can I use my own OpenAI API key with Codex cloud?
An API key runs the CLI, SDK, and IDE extension at published prices, but OpenAI's pricing page says that path has no cloud-based features. You choose between published per-token pricing and the cloud agent.
Where these facts came from
Every mark on this page comes from OpenAI Codex's own documentation and pricing pages, not a third-party roundup: read on October 16, 2026. If something here is out of date or wrong, tell us and we will fix the page.