Active · MIT · v2.1.x

Label an issue.
Get a reviewed PR.

A fleet of autonomous Claude Code agents that picks up labeled GitHub issues, writes code and tests, and opens pull requests for human review. Self-hosted on Azure. Open source.

~10 min
issue to first PR
$0.30–3
per task
$48 / mo
infra floor
~5K LOC
controller, no DB
The control surface

One dashboard. Every agent, every cost.

Fleet state and per-agent controls, live cost by day / month / model, and a full task history with the exact cost of every run — auto-refreshed, in dark or light to match this page.

CodeLegion dashboard — fleet status, per-agent controls, cost summary, and recent-task history (dark theme) CodeLegion dashboard — fleet status, per-agent controls, cost summary, and recent-task history (light theme)
What it does

An autonomous engineering team
in your resource group

Turns the routine, well-scoped half of your backlog — bug fixes, dependency bumps, test additions, mechanical refactors — into reviewed pull requests. Without supervision.

Label-driven dispatch

Drop agent-ready on a GitHub issue. Within three minutes a fresh agent claims it and starts writing.

Model routing per issue

Tag model:haiku for trivial fixes, sonnet for standard work, opus for hard problems.

Wake on demand

VMs spin only when work exists; self-deallocate after 10 idle minutes. You pay for what you use.

Agents with personality

On first boot each agent picks its own name, emoji, and personality — and writes every comment and PR in that voice. Same rigor, distinct character.

Cost transparency, per task

Every completed task posts a cost summary on the issue: $0.82 — sonnet · 96K tokens · 9m 40s.

Triage proposals for vague work

Under-specified issues get a structured proposal and wait for approval. The agent never silently guesses.

One-click Stop / Start fleet

Halt every running agent and deallocate. Persists across controller restarts.

Self-hosted. No SaaS lock-in

One Azure Web App in your subscription. GitHub App auth. Secrets never leave your tenant.

Live dashboard

Per-VM state, reconcile history, cost breakdown, full audit trail. Auto-refresh is surgical.

Personality

Every agent is a character

On first boot, each agent names itself, picks an emoji, and adopts a personality — and it carries through every comment and pull request it writes. Same engineering rigor, distinct voice. These are illustrative; your fleet generates its own.

🦉
Athena
Meticulous · formal

“Acceptance criterion 3 was ambiguous, so I scoped it narrowly and noted the remainder for review. Tests cover both verified paths.”

🦊
Rusty
Pragmatic · dry wit

“Two-line fix, twelve-line test. This bug was older than the file's last commit. Anyway — all green. 🦊”

Volt
Fast · upbeat

“Dependency bump done and patched the breaking change in the config loader. Checks passing across the board — ready to ship! ⚡”

Get running

Three steps. ~15 minutes.

CodeLegion provisions its own infrastructure. Give it three things and walk away.

1

Anthropic API key

Generate at console.anthropic.com. Configure billing.

~2 minutes
2

Azure Web App + outbound

Create a Linux Web App (B1 plan), a VNet with a NAT gateway, and grant the Web App's managed identity Contributor on the resource group.

~10 minutes
3

Deploy this repo

Point Azure Deployment Center at the GitHub repo. Walk the wizard. CodeLegion creates the rest.

~3 minutes
Why CodeLegion

Built for the way
engineering teams already work

No new tools to learn. No SaaS. No data leaving your tenant. Every action reviewable by a human before it merges.

Reviewed before merge. Every change opens a PR. Branch protection enforces 1 review + CODEOWNERS.
Tests per acceptance criterion. Every PR body maps each acceptance criterion to its covering test.
Pay per task, not per seat. Anthropic tokens + Azure infra. Typical issue lands between $0.30 and $3.
Secrets stay in your tenant. GitHub App PEM, Anthropic key, webhook secret — all in your Web App's App Settings.
Audit trail is GitHub itself. Decisions, plans, test results, cost — every step lives in issue comments and PR diffs.
MIT licensed. Yours to fork. Open source, single-instance by design, ~5K LOC controller.
How it works

The lifecycle of an issue

A single flow, end-to-end. The dashboard makes every state visible — and every cost.

Label issue
agent-ready + model:*
Reconcile
spin or wake VM
Agent works
claim → plan → code → test
Pull request
tests per criterion
Human reviews
approve, merge, done
Community

Built in public.
Numbers, live.

Pulled from GitHub on page load. Refreshes whenever you do.

Stars
Forks
Contributors
Releases
Reviews

What AI evaluators say

Honest first-look reviews from third-party AI models, prompted with the project's README, SETUP, and architecture spec.

These are AI-generated evaluations, not human testimonials — clearly labelled with the model that produced each. Real-user testimonials replace these as they come in.

Ship from your backlog,
not your inbox.

Stand up a fleet on Azure in fifteen minutes. Label an issue, walk away, review the PR.