The best background coding agents in 2026, compared

A background coding agent takes a ticket instead of a prompt. You assign it an issue, it checks out the repo somewhere you can't see, writes the change, runs the tests, and comes back with a pull request. If that category is new to you, we wrote a definition post. This one is the shopping guide: the best background coding agents you can put to work in 2026, and how they differ on the details that matter (trigger, execution environment, model choice, pricing).

Product details checked August 2026. Vendors change; check their docs for current specifics.

Full disclosure: we build Sinatra, which leads the list. Every claim about the others comes from their own docs and pricing pages, and we note where a competitor does something we don't.

How we picked the best background coding agents

To make the list, a tool has to actually work in the background. That means four things:

• a trigger surface outside your editor: a tracker issue, a repo label, a chat mention, an API call
• its own execution environment, where it can clone the repo, install dependencies, and run commands
• verification, usually running your tests before it declares the work done
• a pull request as the output, so a human reviews everything before merge

Autocomplete tools and in-editor agents don't qualify, however good they are; Cursor and GitHub appear below only because they now ship cloud agents that meet the bar.

The comparison at a glance

Agent	Trigger	Where code runs	Model choice / BYOK	Pricing model
Sinatra	Assign or @sinatra in Linear; label a GitHub issue	Isolated per-task sandbox, shut down after the run	BYOK: Anthropic, OpenRouter, or a ChatGPT subscription	Free tier; $20/member/mo; no token markup
Devin	Web app, Slack, Linear, Jira, API, Devin Desktop	Cognition's cloud workspaces	Cognition's models; no BYOK	Free; $20 Pro; $200 Max; Teams $80 + $40/seat; enterprise ACUs
Copilot coding agent	Assign the issue to Copilot, agents panel, @copilot on a PR	Ephemeral GitHub Actions environment	GitHub's model list (Claude, GPT, more); no BYOK	Paid Copilot plans from $10/mo; AI credits + Actions minutes
Google Jules	Web app, GitHub, CLI, API	Google-managed VM	Gemini only	Free (15 tasks/day); higher limits via Google AI Pro/Ultra
OpenAI Codex	chatgpt.com/codex, GitHub, Linear, Slack, CLI	OpenAI-managed cloud environments	OpenAI models only	Included in ChatGPT plans; credits or API billing
OpenHands	GitHub label or @openhands, Slack, Jira, API	Your hardware, or OpenHands Cloud	BYOK any provider, or built-in provider at cost	Open source free; free individual cloud tier; enterprise custom
Factory	Factory App, Droid CLI, web and mobile	Local, or Factory-managed cloud machines	BYOK via CLI custom models	$20/$100/$200 usage tiers; enterprise custom
Tembo	@tembo in GitHub, Linear, Slack; Sentry/Datadog events	Tembo cloud VMs, metered by the second	BYOK or ChatGPT OAuth at no inference charge	Usage allowances: $10 trial credit, $60/mo, $200/mo
Cursor cloud agents	Editor, web, iOS, Slack, Linear, GitHub comments, API	Cursor-managed isolated VMs	Cursor's model list, billed at API pricing	Cursor plan + per-token model costs

Sinatra

Sinatra is the Linear-native option on this list. You assign an issue to it the way you'd assign a teammate, or mention @sinatra in a comment; on GitHub you add a label to the issue. It works in an isolated sandbox and runs the test command you configured. It reviews its own PR before asking for yours, and if you leave comments it pushes revisions to the same branch. You run it on your own Anthropic or OpenRouter key, or a ChatGPT subscription you already pay for, with no markup on tokens. Free is 5 tasks a day on your own key; the paid tier is $20 per member per month. It fits teams whose work already lives in Linear and who want the model bill on their own terms. It is not the pick if you need Jira or self-hosting.

Devin

Devin, by Cognition, is the most recognizable brand in the category and by 2026 it is a full platform: cloud sessions, a Slack and Jira and Linear presence, an API, and Devin Desktop, the Windsurf-derived editor that manages local and cloud agents from one view. Sessions run in Cognition's cloud with an interactive IDE you can open and take over. Pricing spans a free tier, $20 Pro and $200 Max individual plans, Teams at $80 per month plus $40 per full seat, and enterprise contracts billed in Agent Compute Units. Model choice is Cognition's own stack; there is no BYOK. If you want one vendor for the whole agent story, services and compliance path included, this is that vendor.

GitHub Copilot coding agent

GitHub's entry (its docs now call it the Copilot cloud agent) is the lowest-friction option if your work already lives in GitHub Issues. Assign an issue to Copilot, or start it from the agents panel, and it works in an ephemeral environment powered by GitHub Actions, then opens a pull request; you can @copilot on the PR to ask for changes. Outbound network access is limited by a firewall with a configurable allowlist. It ships with all paid Copilot plans, from Pro at $10 per month, and consumes AI credits plus Actions minutes as it works. You pick from GitHub's model list (Claude and GPT families among others); you cannot bring your own key.

Google Jules

Google's asynchronous coding agent is called Jules. Hand it a task from the web app, the GitHub integration, a CLI, or an API; it clones your repo into a Google-managed VM, proposes a plan, then executes and opens a PR. It reads an AGENTS.md file in your repo root for context, and it runs on Gemini models only. The free tier is 15 tasks a day with 3 running concurrently; higher limits come with Google AI Pro and Ultra subscriptions, currently for individual Google accounts, with enterprise support still in development. Best for individual developers who want to try the workflow without a purchase order.

OpenAI Codex

Codex is OpenAI's agent, and its cloud mode is the background half: delegate from chatgpt.com/codex, GitHub issues and PRs, Linear, Slack, or the CLI, and it runs in an isolated cloud environment, then hands back a PR. It is included in ChatGPT plans from Free through Enterprise, with usage scaling by tier and credits past the included limits; you can also pay per token with an API key. Model choice is OpenAI-only. If your team already pays for ChatGPT, the marginal cost of trying it is zero.

OpenHands

The open-source pick is OpenHands: a GUI, CLI, and SDK you can run on your own hardware, plus a hosted cloud. On the cloud, you label a GitHub issue openhands or mention @openhands and it attempts a fix and opens a PR; Slack and Jira integrations are part of the hosted offering. Model policy is the loosest here: any provider on your own key, or the built-in provider at cost with no markup. The individual cloud tier is free with a daily conversation cap, and enterprise runs as SaaS or self-hosted in your VPC. Teams that want control over every layer, and will operate some of it themselves, end up here.

Factory

Factory sells "the autonomy stack for enterprise teams," and its agents, called Droids, run from the Factory App, the Droid CLI, and web and mobile, either locally or on Factory-managed cloud machines. The workflow is delegate, review the diff, merge. The CLI supports custom models with your own keys. Pricing is usage-tiered: Pro at $20, Plus at $100 with roughly 5x the usage, Max at $200, and custom Business and Enterprise contracts. The customer list (Blackstone, Adyen, Wipro) tells you the buyer: large organizations standardizing agent workflows.

Tembo

Tembo leans into event-driven triggers more than anyone else here: mention @tembo in GitHub, Linear, or Slack, or wire it to fire from Sentry, PostHog, or Datadog events, and it runs the task on a cloud VM. Billing is a dollar-denominated usage allowance covering model API calls and compute: a $10 one-time credit on the free tier, then $60 and $200 monthly plans. Inference is free through Tembo when you bring your own key or connect ChatGPT via OAuth; you still pay for VM time. If you want a Sentry alert to kick off agent work the way a ticket does, this is the one built around that.

Cursor cloud agents

Cursor's cloud agents extend the popular editor into the background category. Launch one from the desktop app, the web, the iOS app, Slack, Linear, a GitHub or Bitbucket comment, or the API, and it runs in an isolated VM with your repo cloned and dependencies installed, then pushes a branch and a PR. Cloud agents are charged at API pricing for whichever model you select, with a spend limit you set, on top of your Cursor plan. If Cursor is already your editor, its cloud agents are the obvious way to hand off tasks without adopting another tool.

How to choose

Start from where your work lives. If tickets start in Linear, Sinatra is built for exactly that loop; if they start in GitHub Issues, Copilot's agent is one click away; if they start in Slack threads or Sentry alerts, look at Codex, Devin, or Tembo. Then decide who pays for tokens: seat pricing (Sinatra, Copilot plans), platform usage (Devin, Factory, Jules), or pass-through billing on your own key (Sinatra, OpenHands, Tembo, Factory's CLI). Last, decide where code is allowed to run. Only OpenHands lets you keep everything on hardware you control; everyone else runs your repo in their cloud, so read the sandboxing and egress story before you connect a private repo.

Common questions

What makes a coding agent a "background" agent?

It works from a trigger outside your editor, in its own execution environment, and delivers a pull request for human review. You don't watch it type. The contrast is with IDE assistants, which accelerate the work in front of you instead of taking work off your plate; most teams end up running one of each.

Which background coding agent is cheapest?

It depends on volume and whether you already pay for something. Jules has the most generous free tier for individuals at 15 tasks a day, and Codex costs nothing extra if your team has ChatGPT seats. At team scale, BYOK options (Sinatra, OpenHands, Tembo) tend to win because you pay providers directly with no markup, plus a flat per-member or usage fee.

Can I bring my own API key?

With Sinatra, OpenHands, Tembo, and Factory's CLI, yes. Devin, the Copilot coding agent, and Jules have no key option; Codex can bill against your OpenAI API key but stays on OpenAI models, and Cursor bills your chosen model at API pricing through its own account. If key control matters to you, that narrows the list quickly.

If the Linear-to-PR loop is the part you want, assign Sinatra a ticket and see what comes back: start for free, or read the docs for setup. And if you're weighing us against a specific tool above, the alternatives page has the direct comparisons.