Legion for enterprise

A cohort of AI agents, run inside your walls.

Legion is orchestration, not another cloud. It runs locally on your engineers' machines, drives your own model accounts, and every change it makes lands as a normal pull request that passes the exact gates your engineers' code passes.

Enterprise-safe by design

Nothing here is a leap of faith.

No Legion cloud

CLIs plus isolated local git worktrees. No SaaS backend, no Cursor-style vendor in the path. Your repo, history, and worktrees stay on your machine, inference runs on your own Claude and Codex accounts, under your own provider terms.

Your gates, no bypass

Agent output is a normal pull request. It flows through your CI, Sonar, your line ceiling, your duplication threshold, your security scans, and merges only on green. AI gets no special path. Nothing auto-merges; a human approves every change.

Guardrails by construction

Your approved stacks, patterns, and required checks travel as policy, applied identically to every developer and every agent run. Deviations are caught at the gate, not in review three weeks later.

Observable in your stack

Every unit of work emits a legion.span.v1 record, model, tokens, cost, timing, outcome, exported as OpenTelemetry into your Datadog or Grafana. A metered audit trail of what every agent did.

Auditable core

The engine is Apache-2.0 on GitHub. Your security team can review the code, security model, and deployment terms before a pilot, then pin a release and run it from source.

Meets your tools

One shared install works across the harnesses your engineers already use: Claude Code, Codex, Cursor, opencode, DeepSeek Harness, Hermes, and Pi. DeepSeek Harness requires a user-authored dsh profile. No rip-and-replace.

Secure SDLC

Agent code passes the same gate as human code.

Wire your quality gates into Legion and the loop closes cleanly, an AI author earns no shortcut.

01

Legion opens a PR

Task → plan → execute in isolated worktrees → cross-model verify → one pull request.

02

Your gate runs

Sonar static analysis, your 500-line ceiling, 5% duplication threshold, and security scans, on every PR.

03

Green or nothing

Merges only when the gate passes. A pre-PR verify can run those same checks first, so a failing diff never opens a PR.

04

A human approves

Never auto-merged. You stay in control of the line.

AWS Well-Architected

Mapped to the framework you already trust.

Operational excellence
Metered legion.span.v1 traces to OTLP; legion-doctor health-gates the toolchain; self-healing opens (never merges) fix PRs.
Security
Runs locally with your own model accounts and isolated worktrees. Agent pull requests pass your gates before a human approves them.
Reliability
Built for long-running / overnight agents, decompose, isolate, verify, stop at a reviewable PR. Deterministic orchestration, not one unbounded agent.
Performance efficiency
Routing sends each task to the cheapest capable model, escalating only hard work; parallel fan-out across worktrees.
Cost optimization
Spend metered per task on one dollar scale; routing enforces cost ceilings; legion-report shows the bill by model.
Sustainability
Cheapest-capable routing and no idle SaaS layer means less redundant model compute per shipped change.

Deployment

Runs where you run.

today

Local

Legion CLIs on each developer's machine, driving your Claude + Codex accounts; repo and worktrees stay local.

roadmap

Hostable

A centralized / team deployment via API-key-based Claude and Codex.

roadmap

On-network inference

Endpoint-agnostic routing to model endpoints inside your boundary, for teams that need inference on-network.

managed

We build your agent

Opus Aether builds a scoped domain agent in your repos to prove value fast.

The engagement

Start with a scoped pilot.

Legion in one team's repo, against your Sonar gates, with one domain agent for your stack. Success is simple: agent PRs pass your gates and save senior time. Then support and an SLA for the rollout.