The short version

Factory is the more complete product, and it isn't close. It reaches more surfaces, ships managed sandboxes, integrates with the ticketing systems enterprises actually use, publishes benchmark results, and comes with SOC 2, support and an SLA.

Last Light is the more ownable system. It is self-hosted at every tier rather than only at Enterprise, every behaviour is YAML you can read and fork, and the security boundary (egress firewall, per-run token scopes) is yours to set rather than the vendor's to assert.

If you're evaluating both, the question that decides it is rarely a feature: it's whether "on-prem" is a line item you're willing to buy an Enterprise contract for, or a property you need from day one.

Side by side

Last LightFactory droid
Hosting Self-hosted at every tier SaaS; on-prem only on Enterprise
Surfaces GitHub, Slack, dashboard, CLI Terminal, VS Code, JetBrains, Vim, browser, Slack
Model choice BYO: Anthropic, OpenAI, OpenRouter and others Claude, GPT-5.x, Gemini, DeepSeek, Qwen
Ticketing GitHub-issue-centric Jira and Linear first-class
Isolation Docker / micro-VM sandbox + default-deny egress firewall Managed cloud sandboxes (“Droid Computers”)
Customisation Fork any workflow, prompt or persona via the overlay Configurable within the product
Benchmarks None published; ships an eval harness you point at your own repos Published Terminal-Bench leadership
Compliance Whatever you build. You own the boundary SOC 2 Type II, support, SLA
Cost Free / MIT + your model spend + your infra $20 / $100 / $200 per mo + Team/Enterprise

Where each one wins

What Factory does that Last Light doesn't

  • Multi-surface presence. Droids live in the terminal, VS Code, JetBrains, Vim, browser, and Slack. Last Light has no editor integration.
  • Managed cloud sandboxes ("Droid Computers"), zero setup. Last Light expects you to run the sandbox on your host.
  • Ticket-native work. Jira and Linear are first-class. Last Light is GitHub-issue-centric.
  • Broader model menu: Claude, GPT-5.x, Gemini, DeepSeek, Qwen.
  • Published benchmark leadership (state-of-the-art Terminal-Bench results). Last Light publishes none, and hands you an eval harness instead.
  • A commercial product: SOC 2 Type II, support, SLA, managed upgrades, polished UX.
  • Mature multi-agent coordination across Code/Reviewer/Tester/Knowledge/Reliability droids.

What Last Light does that Factory doesn't

  • Truly self-hosted at every tier. Code, prompts, and model traffic stay on your infra with your keys. Factory on-prem is Enterprise-only.
  • Workflows as open config, not a closed product. Read, fork, and override every behaviour via the instance/ overlay.
  • An egress policy you own: a default-deny firewall plus four downscoped, per-run GitHub App token profiles.
  • First-class approval gates. Pause after the architect plan; resume via GitHub comment, Slack, or dashboard.
  • Standing event + cron service: real-time webhooks and scheduled weekly health/security sweeps, plus a weekly per-repo Slack digest of what shipped and what the agent did about it.
  • Per-repo self-service. Each repo can commit a bounded .lastlight/ layer of its own, so teams tune the agent for their code without going through you.
  • Feedback scored against the run. A 👍 or 👎 on the bot's output attaches to the workflow run that produced it, which makes prompt changes measurable.
  • Evals on your own code. The harness replays the real triage, review and build workflows against a mocked GitHub and scores them, with cases built from your own merged PRs.
  • Free and OSS, with no per-seat or per-tier cost.

Which should you pick?

Choose Factory droid when

you want a polished, supported, multi-surface platform your whole team can adopt today, you're fine with a SaaS vendor (or buying Enterprise for on-prem), and you value benchmark-leading models and zero-ops sandboxes over owning the stack.

Choose Last Light when

you need the agent on your own infrastructure with your own keys and a network boundary you control, you want repo-maintenance behaviours as forkable YAML, and you'd rather operate an OSS harness than pay and trust a SaaS, accepting that you own the ops.

Point-in-time as of June 2026. Factory's pricing, model menu and benchmark scores move quickly, so check their pricing page before relying on a figure here.

Compare against something else