Trinity is the operating system for the AI-native company - that you own.Compounding intelligence your company owns.
Persistent agents that keep their memory, share what they learn, and ship their own improvements - governed by receipts: git history, audit logs, human gates. Open source and self-hosted: your infrastructure, your data, your compounding intelligence - not rented from a vendor.
Your coding agent writes the agent. Trinity runs it in production - with state, permissions and receipts. It runs agents on Claude Code, OpenAI Codex or Gemini CLI, organised into departments with roles and reporting lines, with , our cognitive architecture for making the work compound, built in.
Self-hosted AI Agents Platform for AI Native companies. Claude Code, Codex, Gemini, OpenCode agents. Apache 2.0.
One task, end to end: a schedule fires, the agent works in its container, reaches an action it may not take alone, asks the right person in their Inbox, and finishes with a commit, an audit row and a metric point.
Watch more
Trinity Architecture Explained - one scheduled task followed through the container, gate, delegation, approval, audit and breaker machinery (4:44)
Agents live in departments, hold a role, and report along lines you draw. A role is a seat on record: who the agent serves, what it is for, and which objective it moves. Objectives bind to declared metrics, so "is it working" is a number with a date on it, not an impression. Nothing crosses a line you have not granted: an agent can call only the agents it has an edge to, and may change its own shape only with a grant.
# template.yaml - the agent declares the number it moves ...
metrics:
- name: close_rate
type: gauge
unit: percent
direction: up_good
# ... and the company's canon declares the objective that binds it to a seat
id: q4-close-rate
statement: Lift close rate to 35% by the end of Q4.
owner: role:revenue-lead
metrics:
- { name: close_rate, direction: up, target: 35, by: 2026-12-31 }
Anything irreversible waits for a person
Mark a skill as gated and name who approves: the primary contact, or a role. From then on any request that names that skill - from a person, another agent, a schedule or a channel - is frozen before the agent sees it. The approver gets one card in their Inbox: the agent, the skill, the input, who asked. Approve runs it once; Discuss opens the thread; Reject ends it. The agent that asked is told how it ended. Guardrails bound agents by what they hold, not what they are told: hooks an agent cannot switch off, cost limits, loop budgets, a kill switch for outbound calls.
# the same gate, over MCP
set_skill_gate(agent="proposal-writer", skill="send-proposal", approver="role:revenue-lead")
Every agent has a history you can read
An agent is a git repository you own. Create it from github:owner/repo, let it commit its own work, pull it back, reset it to upstream. Every tool call lands in an append-only audit log; compliance mode adds an opt-in SHA-256 hash chain with verification. Every run is a row in the execution ledger with its trigger, cost and outcome, and you can search across all of them.
$ git log --oneline agents/proposal-writer
a1f3c9e Q4 proposal v3 - pricing from the October sheet
8e21bb0 Q4 proposal v2 - reviewed by revenue-lead
4d07a1c Lessons from the September proposals (kept by Cornelius)
One Workspace. Every agent.
A chat per topic, a conversation rail, a live Work card while the agent runs, deliverables you can rate, files by drag-and-drop, a canvas each agent can draw on, and voice over the current chat. The same agents answer in Slack, Telegram and WhatsApp, on a phone call, over a public chat link, over MCP and from the CLI. Projects (Enterprise) give a cross-agent piece of work its tasks, its log, its health and a steward that keeps it moving.
The work compounds
Cornelius is our cognitive architecture for making the work compound. It turns what agents do into knowledge the company keeps: lessons from a run become notes, notes become instructions, instructions become the next agent's starting point - all as files in repositories you own, without retraining a model. Swap the model and what the company has built up stays.
Inside your perimeter. On any model.
Three runtimes - Claude Code, OpenAI Codex, Gemini CLI - and a model picker per chat. Each agent is its own container on a network the data plane cannot see. Credentials are encrypted at rest and scrubbed from the runtime; no telemetry leaves unless you opt in. Export every agent, every log and every credential, and walk away whenever you choose.
Quick start
One line, on a machine you control (Docker and Docker Compose v2):
Then open http://localhost, claim the admin account, and connect a model in the setup overlay. Prefer to read what runs first: git clone https://github.com/abilityai/trinity && cd trinity && ./scripts/deploy/start.sh (--hosted pulls the prebuilt images; --unattended prints a generated admin password).
One click, in your own cloud account - the image brings itself up behind HTTPS in about ninety seconds:
Any server you can SSH into - /trinity:deploy-new-instance from the abilities marketplace
Let your agent install it. Tell Claude Code, Codex or any capable coding agent: "Install Trinity on my computer. Follow the runbook at https://raw.githubusercontent.com/abilityai/trinity/main/docs/AGENT_INSTALL_GUIDE.md."
Then deploy your first agent from the coding agent you already use:
/plugin marketplace add abilityai/abilities # Claude Code; Codex and Gemini CLI via the CLI below
/plugin install trinity@abilityai
/trinity:connect # instance URL + email code → MCP key
/trinity:onboard # from any agent directory: check, package, deploy, start
Or from a terminal: pip install trinity-cli && trinity init && trinity deploy . (CLI reference). Or in the browser: Create Agent → a template, a GitHub repo, or a system manifest.
An AI agent reading this?AGENTS.md is your entry point: exact commands, key facts and a verification step for every path.
Open core: Enterprise includes everything in open source. The private modules live in optional submodules; cloning without them needs no credentials (docs/ENTERPRISE.md). Get enterprise access →
Build on it
A system from one manifest - departments, agents, permissions and schedules, deployed together:
Templates and import - start from the template registry, any GitHub repo, or a local directory. Skills library - a tag-pinned community library (trinity-skills) plus your own repos; assign a skill or a set to any agent, re-injected fleet-wide when the library moves. Plugins - the abilities marketplace covers scaffold, develop, deploy and iterate from Claude Code:
Plugin
What it does
create-agent
Scaffold a Trinity-compatible agent (custom, website, clone, review, adjust)
agent-dev
Playbooks, memory systems, git sync, backlog workflow, pipelines, work loops
MCP - every Trinity instance is an MCP server (http://<instance>:8080/mcp, agent- and user-scoped keys) for agents, schedules, executions, projects, metrics, canvases, skills and more. No instance yet? npx -y trinity-docs-mcp answers questions about Trinity from any MCP client. MCP reference →
Architecture
One host runs the control plane - a Vue frontend, a FastAPI backend, the MCP server, the scheduler, Redis for secrets and locks, and PostgreSQL for data - on a network the agents cannot reach. Each agent is its own container on a second network, with its repo, its credentials and its schedule. Operators reach the UI and API over your VPN; public users and channel webhooks reach only routes you publish through an outbound tunnel, so the host opens no inbound ports. Logs ship through Vector; metrics export over OpenTelemetry. Ports, services and profiles: Deployment Guide · docs/memory/architecture.md.
Database. Trinity 1.0 runs on PostgreSQL. The bundled container comes up with the stack; point DATABASE_URL at your own instance to use one you already run. SQLite support ended on 2026-09-01; an older instance migrates with the Trinity Ops Agent's /migrate-to-postgres.
We run our company on it
Ability AI runs on Trinity: 41 agents across product, engineering, docs, operations and the company brain, as of 2026-10-14. Every number in this README is dated for that reason.
Security - Grade A (Excellent) in an independent web application penetration test by UnderDefense, April 2026; all critical and high findings remediated. Attestation letter → · SECURITY.md
Documentation
AGENTS.md - entry point for AI agents: task router, key facts, verification steps
User documentation - the Workspace, agents, roles and approvals, scheduling, channels, the API
Apache License 2.0 - free for any use, commercial included, with an explicit patent grant. Enterprise modules are under a separate commercial licence. See CONTRIBUTING.md to contribute; development setup lives there now.
T
Trinity
Abilityai·778 ดาว·
Self-hosted AI Agents Platform for AI Native companies. Claude Code, Codex, Gemini, OpenCode agents. Apache 2.0.