Linear coding sessions explained: pricing, setup, and limits

Linear coding sessions turn the tracker into the agent. Since June 2026 you can delegate a Linear issue to the Linear Agent and have it write the code, in a sandbox Linear runs, using Claude Code or Codex, and come back with a pull request on GitHub. There is nothing to install, the pricing is published to the cent, and every team that uses Linear already has it. For anyone evaluating agents for a Linear backlog, it is the baseline the rest of the market now has to beat, and the interesting question is what a month of sessions costs and when a second vendor is still worth having.

Checked in early September 2026 against Linear's docs, pricing page, and changelog, linked throughout.

A disclosure that matters more than usual here. Sinatra is our product, it does the same job, and it is built on Linear and depends on it. This is a comparison of two ways to work, not a complaint about a partner, and there is a whole section below on why most teams reading this should try Linear's version first.

What shipped, and when

Linear introduced the Linear Agent in March 2026 as a workspace agent: it finds related issues, drafts project updates, and pulls requirements out of customer requests. It did not write code.

Coding sessions arrived on June 11, 2026, and changed that. In the changelog's words, "Linear Agent can now write code using Claude Code and Codex," started "by assigning an issue to Linear, or ask it to make a change in a chat, comment, or Slack thread." An August update added configurable environments, browser use, and the pricing that applies today.

Sessions are available on the Basic, Business, and Enterprise plans, per the coding sessions docs, so the free Linear plan is out. Basic is $10 a user a month and Business $16, both billed yearly on Linear's pricing page.

How a Linear coding session runs

Nobody else in this market can match the setup. A Linear owner or admin with GitHub organization owner access grants code access in the GitHub integration settings and enables coding sessions, and after that, per the docs, "any workspace member with a GitHub account linked in Connected Accounts will be able to start a coding session." There is no second vendor, no OAuth app to explain to security, and nothing lands in any repository.

You start one by delegating an issue to the Linear Agent, by mentioning @linear in Slack or Microsoft Teams, or through a triage automation that runs agent behaviors each time an issue arrives in Triage. That last one is a capability nothing else on this page has: incoming bugs can be worked as they land, under rules your team already maintains.

The session runs in a managed sandbox on a workspace-selected agent and model. The docs list Claude Code or Codex as the agent, and a model menu that at the time of writing includes Claude Fable 5, Claude Opus 5, Claude Opus 4.8, Claude Sonnet 5, GPT-5.6 Sol, GPT-5.5, and GPT-5.4, with Opus 4.8 as the default. When the session produces a pull request, "the workflow continues in your Reviews tab," and you can delegate code-review actions back to the agent "to address any left review comments from other teammates."

Repositories have to be reachable through your workspace's GitHub integration. GitLab does not appear in the coding sessions docs, and Linear's GitLab integration page describes merge-request linking only, with no mention of the agent.

What it costs, to the cent

This is the most transparent pricing in the category, so here it is verbatim. From Linear's AI credits page: a coding session's "cost includes model tokens at provider-published rates, with no markup, and sandbox runtime at $0.25 per 20-minute block."

Credits are prepaid and opt-in. You add funds to a workspace-level balance in US dollars before anything is spent; the minimum ad-hoc top-up is $10 and the minimum auto top-up $50. Funds expire twelve months after purchase. If your workspace never adds funds, the docs say, you will not have access to these features and you will never be charged for them. There is no included allowance and no free session.

Because both halves are published, you can estimate a session before running it, which no credit-unit vendor lets you do. Take a session on the default model, Opus 4.8, at Anthropic's list rates of $5 per million input tokens and $25 per million output. Suppose it runs 40 minutes, reads 1.5 million tokens of repository and tool output across its turns, and writes 100,000 tokens of code and reasoning. Those token figures are our estimate for a mid-sized ticket on a real repository, so swap in your own. The sandbox is two blocks, $0.50. The tokens are $7.50 in and $2.50 out. Call it $10.50, and less with prompt caching, which the provider rates include at a lower price for cached input. A tighter ticket on Sonnet 5 at $2 and $10 per million would land nearer $3.

Forty sessions a month at that size is roughly $400 of credits, with no per-member charge for the agent, since you are already paying for Linear seats. That is the number to compare against everything else.

Where a separate agent still makes sense

Most teams evaluating this should use Linear's sessions first. They are already in the workspace, the pricing is legible, and every vendor you do not add is a security review you do not run. The answer changes in a few specific situations.

Your backlog is partly in GitHub Issues. Linear coding sessions have no GitHub-issue trigger; GitHub is where the pull request lands, not where work begins. A team with some repositories whose issues never made it into Linear has to pick a front door, and a separate agent that starts from both trackers avoids the choice.

Your developers already pay for Claude or ChatGPT. Linear sessions draw on the prepaid balance, at provider rates, with no way to point them at a subscription someone already holds. If five people on the team have Claude Max or ChatGPT Pro plans, that is several hundred dollars a month of model capacity that Linear's sessions cannot use, and an agent that accepts those subscriptions as credentials makes the model line zero.

You want to try it before topping up. There is no free session. Some teams want to see a pull request on their own codebase before putting a card on file, and a daily free allowance elsewhere covers that.

The full comparison has the table, and it includes the rows where Linear wins, which are the setup row and the pricing row. The roundup of coding agents for Linear covers the other agents that plug into the same workspace.

Where Sinatra sits next to it

Sinatra starts from the same Linear assignment and also from the Sinatra label on a GitHub issue. It runs on the credential you connect, an Anthropic or OpenRouter key or a Claude or ChatGPT subscription each member links for their own tasks, so the model bill is your provider's or nothing. Per-repository configuration can live in a committed .sinatrarc, reviewed like any other change. A changes-requested review is the instruction, so nobody has to delegate a follow-up action. The seat is $20 per workspace member and the free tier is 5 tasks a day with your own API key or subscription, with no card. Against the estimate above, a team of four running forty sessions a month pays $80 in seats plus whatever their provider charges for the same tokens, or $80 flat if subscriptions cover the models. Do that arithmetic with your own numbers; for a team running a few sessions a week on no subscriptions, Linear's version is cheaper and you should use it.

Common questions

Do Linear coding sessions mark up model tokens?

No. The docs say tokens are charged at the provider's published rates with no markup, and sandbox runtime is $0.25 per 20-minute block, both drawn from a prepaid balance.

Are coding sessions on Linear's free plan?

No. The docs list Basic, Business, and Enterprise. Basic is $10 a user a month billed yearly.

Can I use my own Anthropic or OpenAI key?

Nothing in Linear's docs describes a bring-your-own-key path. Sessions run on a workspace-selected model billed from the credit balance.

Can a coding session start from a GitHub issue?

No. Sessions start by delegating a Linear issue, mentioning @linear in Slack or Teams, or through a triage automation. The pull request lands in GitHub; the work starts in Linear.

Running both is the easiest comparison in this market to set up, since delegating to the Linear Agent starts one and assigning to Sinatra starts the other, on separate branches, with nothing merged by either. Give each a project for a week and compare the pull requests and the two costs. If the credit draw is smaller, use Linear's. If you want the other trigger or the subscription, start free and put yours behind it.