AI Coding AgentsDeveloper Tools

Talyn

Checked November 20, 2026. Free with 10 box-hours a month and unlimited local worktree sessions; Pro $40/month (200 box-hours, or $400/year); Max $100/month (1,000 box-hours); team seats $40 each. You bring your own Claude or ChatGPT subscription, so model usage is not included. Extra sandbox time is billed per second from prepaid credits. The source code is MIT-licensed. Check official pricing →

What Talyn Does

Coding agents are good at producing a diff. The slow part comes after: opening the pull request, waiting for CI, fixing the test that broke on Linux but not on your Mac, resolving a conflict with someone else’s merge, and answering review comments. Talyn is built around that second half. Each task gets its own cloud sandbox and git branch, Claude Code or Codex does the work there, and Talyn keeps the resulting pull request moving until it is green and, if you ask, merged.

It is a desktop app for macOS, Windows, and Linux, plus a browser version at app.talyn.dev, built by an independent developer and published as open source under the MIT license. As of November 20, 2026 it is in public beta, it works only with GitHub, and new releases ship several times a day.

How It Works

Talyn has three building blocks.

Sessions. A session is one task: a branch, a sandbox started from an image of your repository, a coding agent, and a terminal. You describe the task, read the diff as the agent works, open a terminal or start the app inside the sandbox, and press one button to open the pull request. Sessions are isolated from each other, so two agents never edit the same files. Close your laptop and the sandbox keeps going; an idle sandbox sleeps after 10 minutes and costs nothing while asleep.

Threads. A thread is one conversation that can start sessions across several repositories. Give it a goal such as “show usage in the app,” and it opens a session for the API, one for the web front end, and one for the docs, follows them, and stops to ask you when a session needs a decision.

Pull requests. A dashboard lists the PRs you own and the reviews you owe, worst first. From there you can send an agent to fix failing checks, a merge conflict, or review comments on the same branch, run a multi-reviewer code review (findings stay in the app unless you post them), and queue a ready PR so it merges when it turns green.

Around those, Talyn adds automations: GitHub rules (“when checks fail on a dependency bump, send an agent”), Slack triggers (a mention or reaction starts a session), and scheduled prompts on a cron schedule. It also runs your SKILL.md playbooks on pull requests, connects to MCP servers such as Linear or Sentry, and has a public API and hosted MCP server for starting sessions from your own scripts.

Supported Agents and Models

Talyn runs Claude Code and Codex. Each signs in with your own Claude or ChatGPT subscription through Anthropic’s or OpenAI’s own login page, so there is no API key to paste and no token bill from Talyn. You can switch agents during a conversation. Which Claude or OpenAI models a session can use, and how much, depends on your plan with Anthropic or OpenAI; Talyn’s repository shows GPT-6.2 Sol added for Codex runs on November 17, 2026, the day OpenAI released it.

In the desktop app, a session can also run locally in a git worktree on your own machine with your own tools. Local sessions use no sandbox and are free and unlimited on every plan. You can also move a Claude Code conversation from your computer into a cloud sandbox and carry on there.

A Practical Workflow

Say you have two unrelated jobs before a release: add CSV export to a reporting page, and fix a flaky login test.

  1. Start two sessions. Give the CSV export to Claude Code and the flaky test to Codex. Each gets its own branch and sandbox, so neither can trample the other’s files.
  2. Watch or walk away. Read the diffs as they form, or close the app. When the Codex session finishes, you run the test suite in its sandbox terminal yourself before trusting its claim that the flake is gone.
  3. Open the pull requests. One click each. CI runs as it normally would.
  4. Let Talyn handle the follow-up. CI fails on the CSV branch because of a lint rule, and a teammate leaves a review comment asking for a header row. You send a fix run; the agent pushes to the same branch. Meanwhile the login-test PR falls behind main and picks up a conflict, and a second fix run resolves it.
  5. Review and merge. You read the final diffs and approve them. If you flagged a PR for the merge queue, Talyn merges it through GitHub once checks pass, and your branch protection and required reviews still apply.

What Talyn does not do is decide whether the change is right. A green CI run only proves your tests pass, and an agent that “fixes” a failing check can do it by weakening the test. Reading the final diff is still your job.

Pricing

You pay Talyn for sandbox time, called box-hours, and pay Anthropic or OpenAI separately through your existing subscription.

PlanPriceIncludedNotes
Free$010 box-hours a month; up to 2 boxes at onceSmall and standard boxes; the full app, including merge queue and automations
Pro$40/month, or $400/year200 box-hours; up to 100 boxes at onceAll box sizes; auto-keep and auto-review on every new PR; credits for overage
Max$100/month (monthly only)1,000 box-hoursHalf the per-hour cost of Pro
Team$40 per seatPro for each seat, 200 box-hours eachNo shared workspaces, SSO, or admin controls yet

A box-hour is one standard box (4 vCPU, 4 GiB) running for an hour. Above your allowance, prepaid credits cover overage at $0.11 to $0.55 an hour depending on box size, billed per running second. Prices are from talyn.dev, checked November 20, 2026.

Remember the other bill. Ten parallel sessions of Claude Code or Codex consume your Claude or ChatGPT plan’s limits ten times as fast as one, and a fix run that loops on a stubborn test spends usage too.

Security and Privacy

Talyn publishes a clear security page, and the source is public, so you can check it.

  • Credentials stay outside the sandbox. Agent calls to Anthropic or OpenAI, and calls to MCP servers, pass through a proxy that adds your credential. An agent tricked by a hostile file into looking for keys finds none.
  • Isolation. Each session runs in its own microVM. A sandbox cannot reach your computer or a private network, and it is deleted when you archive the session.
  • Network access. Sessions have internet access by default so agents can install tools and read docs. You can turn that off per session or as a default, leaving only the repository, package registries, and the agent’s API. Fix runs, code reviews, and scheduled automations are limited to the repository and agent API unless you allow more.
  • GitHub permissions. The GitHub App sees only the repositories you install it on and needs write access to code so it can push fixes. It does not need organization admin rights.
  • What it stores. Your GitHub identity and PR metadata, encrypted integration tokens (AES-256-GCM), session transcripts, and unpushed work saved after each turn so a lost sandbox can be restored.
  • Untrusted input. Talyn says it marks PR text, review comments, Slack messages, and repository files as untrusted for the agent, and that its sandbox limits do not rely on the agent obeying that label. That is the right approach to tool-output prompt injection: a review comment written by an outsider should not be able to instruct an agent with write access.

Your code is also processed by Anthropic or OpenAI under your plan’s terms. Read Talyn’s privacy policy before connecting a private repository.

Strengths

  • Turns a pile of agent diffs into pull requests that actually reach a mergeable state
  • Runs Claude Code and Codex side by side on the subscriptions you already pay for
  • Clean isolation: one branch and one microVM per task, with credentials kept outside
  • Free local worktree sessions and a usable free tier
  • Open source, so the security claims can be inspected

Limitations

  • Early and small. Public beta, built and run by an independent developer, with releases several times a day. Expect rough edges and plan for the project changing direction or stopping.
  • GitHub only. No GitLab or Bitbucket.
  • Two agents. Claude Code and Codex only; no Gemini, Cursor, or OpenCode agents.
  • Team features are thin. No shared workspaces, SSO, or admin controls yet. Windows installers are not code-signed, so SmartScreen warns on install.
  • Your plan limits still apply. Parallel agents burn through Claude and ChatGPT allowances quickly.
  • Automation can hide judgment calls. Auto-fixing checks and auto-merging are convenient, but they can paper over a weakened test or a wrong fix. Keep merges opt-in until you trust it on your codebase.

Continue learning

Explore related guides, tools, workflows, and prompts that help you go deeper into this topic.

See how this tool fits into a workflow

Browse step-by-step AI workflows that use ChatGPT, Claude, Gemini, and other tools.

Frequently Asked Questions

What is Talyn?

Talyn is a desktop and web app that gives each coding task its own cloud sandbox and git branch, hands the task to Claude Code or Codex, and then takes the result through review: it opens the pull request and can send an agent to fix failing CI checks, merge conflicts, and review comments until the PR is ready to merge. It is in public beta and works with GitHub only.

Do I need an API key to use Talyn?

No. Claude Code and Codex sign in with the Claude or ChatGPT subscription you already have, and Talyn charges only for the sandboxes. Those runs draw on your plan's usage limits, so several agents running in parallel will use them up faster.

Will Talyn merge code without asking me?

Only pull requests you flag. Keeping a PR mergeable and adding it to Talyn's merge queue are opt-in, one pull request at a time, unless you turn on a Pro or Max workspace setting that flags every new PR. Merges go through GitHub's normal merge API, so your branch protection rules and required reviews still apply.

Does Talyn work with GitLab or Bitbucket?

No. Talyn installs as a GitHub App on the repositories you choose, and its maker says GitLab and Bitbucket support is not planned soon.

How is Talyn different from Codex cloud or Claude Code on the web?

Those run one vendor's agent in that vendor's cloud. Talyn runs either Claude Code or Codex in its own sandboxes, lets one conversation start sessions across several repositories, and adds pull-request work on top: a dashboard of your PRs and review requests, fix runs for failing checks, a merge queue, and automations triggered by GitHub, Slack, or a schedule.

Is Talyn safe to connect to my repositories?

Talyn's documented design is reasonable: the GitHub App sees only repositories you pick, each session runs in its own microVM, and your agent sign-ins and MCP keys are added by a proxy outside the sandbox so an agent cannot read them. It does need write access to code, stores session transcripts and unpushed work, and sandboxes have internet access by default. It is a young project from an independent developer, so start with a non-critical repository.

Last updated: