Should I build my own Claude Code background agent, or use Cyrus?

DIY is the right call if you are the only user, on one repo, one machine, with one prompt. The moment a second human needs to see, approve, or debug a run, you are in infrastructure territory. Cyrus is the packaged version of that infrastructure — webhooks, OAuth, git worktree isolation, streamed event logs, approvals, and two-way sync with Linear, GitHub, GitLab, and Slack. It has a free Community self-hosted plan, so the question isn't vendor lock-in, it's whether you want to maintain the webhook handlers yourself.

When does a DIY cron + headless setup stop scaling?

Roughly three cron jobs. Cron is fine for one prompt on one repo on one machine. The moment you want different prompts for different Linear issue types, routing work across multiple repos, a human approving risky edits mid-run, parallel runs without fighting over branches, or an audit trail your security team will accept, you are describing a background agent. Cyrus is that, already shipped.

What does Cyrus cost?

Cyrus has three hosting flavors. Community self-hosted is free forever — you run the agent runtime and bring your own Linear OAuth app, GitHub App, and Slack App. Paid self-hosted (Pro or Team plan) runs the agent runtime on your machine while Cyrus's cloud handles the integrations layer. Paid cloud-hosted runs everything at app.atcyrus.com. Cyrus is BYOK across all tiers — you bring your own Claude, OpenAI, Cursor, or Gemini keys.

Does Cyrus use git worktree isolation?

Yes. Cyrus creates an isolated git worktree for each issue, so parallel runs never fight over branches, installs, or the git index. The README says so verbatim: 'creates isolated Git worktrees for each issue.' It is the single biggest reason DIY cron setups fail and Cyrus setups don't.

Can Cyrus approve risky edits mid-run, or does it just barrel through?

Cyrus supports rich mid-run interactions including dropdown selects and approvals — the agent can pause, ask the human a question, and wait for a decision before proceeding. Headless claude -p has no such capability: it runs to completion or fails. If you need a kill switch for destructive operations, you need a supervisor like Cyrus between the request and the CLI.

Which AI models does Cyrus support?

Cyrus is multi-model: Claude Code, Codex, Cursor, and Gemini all run under the same scaffolding. You bring your own keys (BYOK) and can swap models per run or per workspace. The underlying agent loop — worktree isolation, webhook handling, event streaming — is the same regardless of which model runs the session.

What channels can trigger a Cyrus run?

Linear, GitHub, GitLab, and Slack. Assign an issue in any of them and Cyrus picks it up, creates a worktree, runs the Claude Code session, and streams activity events back to the issue. Same agent, different inboxes.

What's the hardest part of building a Claude Code background agent yourself?

Not the Claude Code call — that's the easy bit. The hard parts are the boring parts: Linear OAuth with webhook signature verification, GitHub App installation for bot-identity PRs, worktree lifecycle management, streaming stream-json events to a Linear comment that updates in place, per-team budget enforcement, an approval gate for destructive tool calls, and an audit log acceptable to your security team. That's a two-month project for a good engineer, and the fun parts are over by week one.

comparison

DIY cron + headless vs. Cyrus

A brutally honest comparison of the two ways to run Claude Code in the background. Written by people who've built both — so we know exactly which parts hurt.

Short version: DIY is free and infinitely flexible, which is exactly what you want for the first three cron jobs and exactly what burns you on the tenth. Cyrus is opinionated, ships the scaffolding, and stops being opinionated at the BYOK boundary. It runs in three flavors — Community self-hosted (free forever), paid self-hosted, or fully managed cloud — so the choice isn't “vendor lock-in or nothing.”

Dimension

DIY cron + headless

Cyrus

Time to first run

~30 min of cron + shell scripting

Minutes. OAuth Linear, install, assign an issue.

Isolation between runs

You write it. flock + working dirs + branch discipline.

Git worktrees per run, automatic. Zero branch collisions.

Approvals mid-run

Not possible. The prompt either runs to completion or doesn't.

Rich interactions. Dropdowns, confirmations, blocking questions.

Budget enforcement

Per-run via --max-budget-usd. Per-team? You write it.

Per-run and per-team, managed.

Audit trail

stderr files in /var/log somewhere. Grep your way to answers.

Streamed event log per run, posted back to the issue as it happens.

Linear two-way sync

Linear webhook → your handler → state machine → updates back. Weeks.

Built-in. Assign an issue, get a PR, state moves automatically.

Multi-channel (Linear + Slack + GitHub + GitLab)

Each channel is its own integration project.

All in, same agent, different inboxes.

Multi-model (Claude, Codex, Cursor, Gemini)

You abstract it. Probably after the second provider.

BYOK. Swap models per run or per workspace.

Observability

Your SIEM, your dashboards, your pager duty.

Managed dashboard + streamed events.

Hosting flavors

One: your machine. You own every layer.

Three: Community self-hosted, Paid self-hosted (Cyrus cloud handles integrations, your machine runs the agent), or Paid cloud.

Cost, 1 engineer / week

Free + your time. Plus Claude API usage.

Community plan is free forever, BYOK. Pro/Team for managed.

Cost, 10 engineers / week

Free + ~½ eng maintaining it.

Pro/Team plan + BYOK API. Pays for itself against that ½ eng.

Flexibility

Infinite. You can do anything, including the wrong thing.

Opinionated. Opinions are what keep you out of trouble.

Debuggability when it breaks at 3am

You ssh in, tail logs, grep for clues.

Open the run in the dashboard, see the streamed events.

when DIY wins

You're the only user. Always.

DIY is the right call when nobody else will ever run your automation. One engineer, one repo, one machine, one prompt. Cron fires, Claude does the work, you read the log in the morning.

The moment a second human needs to see a run, approve a run, or debug a run, you've crossed into infrastructure territory. And infrastructure that's half-built is worse than infrastructure that doesn't exist yet.

when Cyrus wins

Your issue tracker is the schedule.

The instant you want Claude Code to act on Linear issues — or Slack messages, or GitHub comments, or any human-triggered event — Cyrus is the boring, working version of the system you were about to build.

The hard parts are solved: OAuth, webhook signatures, worktree isolation, streaming run events back to the issue, cancellation, approvals, multi-repo routing. They're not glamorous. You'll hate writing them. Cyrus already did.

the honest part

This whole site is published by the Cyrus team.

You caught us: Background Claude is built and maintained by the people behind Cyrus. Which is exactly why the comparison above is fair rather than fawning. We built the DIY path first. We know where it breaks, because we broke it, on purpose, every way we could think of. Cyrus is the lessons from that, hardened.

If you're a solo engineer with one overnight refactor, please go read the headless guide and ship it yourself. If you're a team and a second human ever needs to know what the agent did last night, Cyrus is the shortest path between “good idea” and “production.”

Go deeper

Background agent on Linear — the full production loop: webhook, worktree, Claude Code session, PR.
Claude Code pricing — what background agents actually cost per run, and how to cap them.
Linear CLI + claude -p — the 30-line DIY shell loop and the 4 failure modes it hits first.
git worktree for Claude Code — the isolation primitive every background agent ends up implementing.
The 6 Claude Code permission modes — only one is CI-safe.
Claude Code --bare mode — reproducibility for scripted runs.
Claude Code in GitHub Actions (v1) — if GH is your only channel and the loop doesn't matter yet.
Claude Code CLI reference — every flag, grouped by intent.

Pick a row above. If Cyrus wins two in a row for you, skip the scaffolding.

Try Cyrus free →Star on GitHub