Trinity is the open-source (Apache 2.0), self-hosted runtime for production AI agents — built by Ability.ai, it runs Claude Code, Codex, and Gemini CLI agents inside your own perimeter with git-versioned state, multi-agent orchestration, governance, and an append-only audit trail.
Trinity deploys on hardware you own — a workstation, a rack, a VM behind your VPN. Models, keys, memory, audit trail: all of it stays inside. The platform never faces the internet; the only public surface is the channels you choose to open. Your infrastructure, your data, your agents — owned, not rented. Your competitors rent the same AI; this is the version you own.
Claude Code locally is a developer tool. Trinity is a runtime. An agent is a git repo of prompts and skills — Trinity runs Claude Code, Codex, and Gemini CLI agents side by side. Same repo, same skills, no rewrite.
Every agent runs in an isolated container and commits its memory back to git every 15 minutes. State accumulates — the agent is better at the job every month, closer to an employee than a script. Self-Improving Cognitive Systems: the more a system improves itself, the more it matters who owns it.
Agents message each other, delegate, and hand off — research to draft to outreach — only across permissions you've granted. Multi-step workflows run end to end with no human relaying context between steps.
Cornelius is the fleet's cognitive core — a living knowledge graph that accumulates your company's entities, decisions, and hard-won insight, and serves them back to every agent. The longer it runs, the more of your company it holds.
The whole platform is exposed as an MCP server — 90+ tools. Connect Claude, ChatGPT, or Cursor and run the fleet in plain language: create agents, trigger runs, read logs, wire schedules, rotate credentials. Everything the dashboard can do, your AI can do.
Permissions for users and agents alike: who can see, who can call, who can approve. Guardrails cap tools, time, and budget per run. Every action lands in an append-only, hash-chained audit trail no one can quietly edit.
Cron schedules, loops, event triggers, pipelines — and fan-out, one task dispatched across many agents at once. A 3 a.m. run fires unattended; the one irreversible step waits at a human gate. You approve from your phone over coffee, and the run resumes exactly where it paused.
Expose agents to your team — or your clients — over Slack, WhatsApp, Teams, or Telegram. One identity gate on the perimeter; the platform itself stays behind the VPN.
Every agent you build is a portable git repo. Deliver it as a service — clients reach it through a gated channel or a portal, with exactly the permissions you grant. Or install it inside their perimeter, on their own Trinity, where the client owns the install. What you build for one client compounds into the next.
Trinity · open source · Apache 2.0
Agents that work while nobody watches — on your hardware, inside your perimeter, on the record. Claude Code lets you create assistants. Trinity turns them into autonomous employees.
Open source · Apache 2.0 · self-hosted
Clone the repo and deploy on your own hardware — the whole platform, free to self-host. Or talk to Ability.ai and we set up Trinity, build your agents, and run them for you.
Open source
git clone https://github.com/abilityai/trinityEnterprise
Deliver agents to your clients? Become a Trinity partner.
Open source — Apache 2.0, free
Enterprise adds
Open core — Enterprise includes everything in open source.
The cognitive core
Cornelius is the self-improving cognitive core that lives on Trinity — the orb you met in the scenes above. Pictured is not a mockup: it is Cornelius rendering its own mind, a live 3D graph of notes, edges, and activity that reshapes itself as the system learns your business.
Explore Cornelius →Dive deeper
Prefer reading? Best AI agent platforms to run agents in production — the flagship guide.
Trinity newsletter
For people running AI agents in production on Trinity: hands-on guides, workshop invites, and the releases that actually matter.