Trinity

Community

Sovereign AI Agents Platform supporting Claude Code, Codex, Gemini agents. Apache 2.0.

How to install ↓ Compare GitHub source repository ↗

Category: Dev Tools #95 of 438
Stars: ★ 637 Very popular
Transport: Streamable HTTP
Runtime: Docker · Python 3.11+
Credentials: API key / credential required
License: Apache-2.0
Last commit: 8 days ago
Tools: 77

62FMRS · C

Trinity is a feature-rich self-hosted AI agent orchestration platform. It provides extensive MCP tools for agent management, scheduling, communication, monitoring, and governance. Suitable for production use, but requires Docker and self-hosting infrastructure. MCP integration is via streamable-http and requires authentication.

Strongest · Maintenance 15/20 Weakest · Reliability 9/20

Reliability

9/20

Security and permissions

12/20

Maintenance

15/20

Documentation

14/20

Setup experience

12/20

Why each score

Reliability 9/20: Evidence: README shows a complete compose architecture (FastAPI backend, Vue frontend, MCP server under src/mcp-server with 90+ claimed tools, scheduler, Redis/SQLite/Vector); deployment scripts (install.sh, start.sh, quickstart.sh); a testing guide and hygiene skills are mentioned. Deductions: static review cannot execute anything; no CI workflow files or committed test suite paths are provided as verifiable evidence; the MCP tool list is only asserted in the README with no actual tool schemas; 272 open issues and a multi-container dependency set (Docker, Redis, Postgres, Vector, OTel) make clean startup and controllable failure modes uncertain.
Security and permissions 12/20: Evidence: README explicitly describes Docker container isolation per agent, encrypted credential storage (.credentials.enc), a 4-tier RBAC hierarchy, append-only audit trail, HMAC webhook signatures, x402 payment confirmation, verified-email allow-lists, configurable guardrails, and an independent UnderDefense pentest (Grade A); SECURITY.md is referenced as a disclosure channel. Deductions: none of these claims are backed by source-level verification in this static review (actual RBAC enforcement, audit immutability, encryption details); install examples use localhost and Bearer placeholders with no real tokens found; dangerous operations (remote deploy, payments, ephemeral SSH) lack explicit confirmation-step documentation in the README.
Maintenance 15/20: Evidence: Apache-2.0 license; README indicates an active release cadence (v0.8.5 dated 2026-07-26) with versioned release notes, an automated CLI publishing path to PyPI/Homebrew, CONTRIBUTING.md, SECURITY.md, and GitHub Issues/Discussions channels. Deductions: 430 stars against 272 open issues is a high ratio; commit/release frequency cannot be verified from the untrusted README; no concrete dependency-update or security-advisory process evidence is supplied.
Documentation 14/20: Evidence: README is a layered docs hub: quick start, manual install, agent templates, multi-agent YAML examples, MCP client JSON config, 90+ tools grouped by category, CLI reference, deployment guide, PostgreSQL setup, known issues, testing guide, security attestation, and an AGENTS.md router for AI agents. Deductions: MCP tools table lacks parameter/input schemas, error codes, and rate limits; the 90+ tool count is claimed but no actual tool definition files were shown; known-issues doc is linked but its content is not evidenced; auth/install steps rely on README assertions that static review cannot confirm.
Setup experience 12/20: Evidence: README offers multiple install paths: one-line curl script, manual Docker steps, interactive quickstart.sh, CLI via pip/brew, Claude Code plugin marketplace, and a ready-to-paste MCP client JSON example (streamable-http at localhost:8080/mcp with Bearer auth); prerequisites are only Docker plus an API key. Deductions: no CI workflow or committed tests are provided as verified execution evidence, so the static calibration caps setup at 15; the multi-container dependency set (Docker Compose, Redis, PostgreSQL, Vector, OTel) and private optional submodules make installation potentially fragile; no Windows/other-platform compatibility notes; end-to-end verification steps after MCP connection are thin.

Static review · not runListed 2026-08-07

Read the FMRS scoring method →

Fit and risk

What it can accessRuns commands or codeUses the networkConnects to a database

Best for

Organizations needing self-hosted AI agent deployment with data residency and audit.
Development teams wanting to run agents with Claude Code, Codex, or Gemini.
Production-grade agent orchestration with isolation, scheduling, and observability.

Not for

Users needing zero-configuration hosted solutions (requires Docker and self-hosting).
Those seeking a lightweight agent framework without Docker.
Teams not interested in managing the platform themselves.

Required permissions

MCP server requires an API key (Bearer token) for authentication.
Agents themselves run in Docker containers and may require network access to APIs and external endpoints.
The platform stores credentials (encrypted) and Git data.
Admins and users control access through role hierarchy (admin, creator, operator, user).

Risks and side effects

Self-hosting requires infrastructure management; misconfigurations could lead to security issues.
Agent-run code may be untrusted; Docker isolation is essential.
Credentials are stored via Git and encrypted, but need proper protection.
Webhooks may expose public endpoints, mitigated with HMAC signatures.

Setup

Before you start

Runtime：Docker · Python 3.11+

SECRET_KEY requiredsecret JWT signing key, generate with openssl rand -hex 32.

ADMIN_PASSWORD requiredsecret Admin login password, minimum 8 characters.

ANTHROPIC_API_KEY optionalsecret Anthropic API key for Claude-powered agents, can also be set via the Settings UI.

GITHUB_PAT optionalsecret GitHub personal access token for cloning private template repos.

Other optional settings (4)

DATABASE_URL optional Optional PostgreSQL connection string, recommended for production; defaults to SQLite if unset.

OTEL_ENABLED optional Enable OpenTelemetry metrics export, default false.

EMAIL_PROVIDER optional Email provider for verification codes: console (dev), smtp, sendgrid, or resend.

EXTRA_CORS_ORIGINS optional Additional allowed CORS origins.

Clone the repository: git clone https://github.com/abilityai/trinity.git && cd trinity
Copy .env.example to .env and set SECRET_KEY (openssl rand -hex 32) and ADMIN_PASSWORD.
Run ./scripts/deploy/build-base-image.sh to build the base image.
Run ./scripts/deploy/start.sh to start all services.
Open http://localhost to complete the setup wizard, set admin password, and configure API keys.
Configure MCP clients with 'mcpServers' pointing to http://localhost:8080/mcp, using 'Authorization: Bearer YOUR_API_KEY' header.

claude_desktop_config.json

{
  "mcpServers": {
    "trinity": {
      "type": "streamable-http",
      "url": "http://localhost:8080/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

Shown for Claude Desktop. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.

.vscode/mcp.json

{
  "servers": {
    "trinity": {
      "type": "streamable-http",
      "url": "http://localhost:8080/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal

claude mcp add --transport http trinity http://localhost:8080/mcp --header 'Authorization: Bearer YOUR_API_KEY'

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

After deployment, connect your MCP client to http://localhost:8080/mcp; the tool list should include list_agents, chat_with_agent and others. Calling list_agents and getting back an agent list confirms the connection works.

Troubleshooting

Ensure Docker daemon is running and you have permissions.
Check backend logs: docker compose logs -f backend.
Verify SECRET_KEY and ADMIN_PASSWORD are set in .env.
Ensure ports 80, 8000, 8080 are not in use.
Test MCP connection: curl http://localhost:8080/mcp.

Things to try

Once connected, you can ask your AI assistant things like:

List all agents and their status
Send a message to my-agent asking what it can do
Create a daily 9am schedule for my-agent
Show me the recent container logs for my-agent

Tools 77

list_agents read-only

List all agents with status.

get_agent read-only

Get detailed agent information.

get_agent_info read-only

Get agent template metadata (capabilities, commands, etc.).

create_agent writes

Create a new agent from template.

deploy_local_agent writes

Package and deploy a local agent directory.

start_agent writes

Start an agent container.

stop_agent writes

Stop an agent container.

rename_agent writes

Rename an agent.

Show 69 more tools

delete_agent destructive

Delete an agent.

initialize_github_sync writes

Initialize GitHub sync.

set_agent_github_pat writes

Set per-agent GitHub PAT.

chat_with_agent writes

Send a message and get response (supports parallel and async modes).

get_chat_history read-only

Retrieve conversation history.

get_agent_logs read-only

View container logs.

fan_out writes

Dispatch N parallel tasks to an agent and collect results.

send_message writes

Proactive user message by verified email (agent-initiated).

send_group_message writes

Proactive group messaging.

list_channel_groups read-only

List channel groups.

deploy_system writes

Deploy multi-agent system from YAML manifest.

list_systems read-only

List deployed systems with agent counts.

restart_system writes

Restart all agents in a system.

get_system_manifest read-only

Export system configuration as YAML.

list_templates read-only

List available templates.

inject_credentials writes

Inject credential files directly to agent.

get_credential_status read-only

Check credential files.

get_agent_ssh_access writes

Generate ephemeral SSH credentials for direct terminal access.

list_schedules read-only

List all schedules.

create_schedule writes

Create a new schedule.

update_schedule writes

Update a schedule.

delete_schedule destructive

Delete a schedule.

enable_schedule writes

Enable a schedule.

disable_schedule writes

Disable a schedule.

trigger_schedule writes

Trigger a schedule manually.

list_executions read-only

List recent executions.

poll_async_result read-only

Poll for async results.

get_agent_activity read-only

Get agent activity summaries.

list_tags read-only

List all agent tags.

get_tag read-only

Get a tag.

set_tag writes

Set a tag.

add_agent_tag writes

Add a tag to an agent.

remove_agent_tag writes

Remove a tag from an agent.

list_subscriptions read-only

List subscriptions.

register_subscription writes

assign_subscription writes

Assign a subscription to an agent.

unassign_subscription writes

Unassign a subscription from an agent.

list_skills read-only

List skills.

create_skill writes

Create a skill.

update_skill writes

Update a skill.

delete_skill destructive

Delete a skill.

assign_skill writes

Assign a skill to an agent.

get_fleet_health read-only

Get fleet health.

get_agent_health read-only

Get agent health details.

trigger_health_check writes

Trigger health checks.

configure_payments writes

Configure Nevermined x402 payments.

toggle_payment writes

Toggle payment.

view_payment_history read-only

View payment history.

send_notification writes

Send structured notifications from agents to platform.

emit_event writes

Emit events.

subscribe_to_events writes

Subscribe to agent events.

list_event_subscriptions read-only

List event subscriptions.

delete_event_subscription destructive

Delete an event subscription.

git_status read-only

Get Git status.

git_sync writes

Sync Git.

git_log read-only

Get Git log.

git_pull writes

Pull Git changes.

git_sync_state read-only

Get sync state.

git_recovery_reset destructive

Recovery reset.

list_operator_queue read-only

Read human-in-the-loop approval items.

respond_to_approval writes

Respond to approval items.

run_loop writes

Run bounded sequential task loops.

monitor_loop read-only

Monitor loops.

stop_loop writes

Stop loops.

publish_report writes

Publish structured agent reports to dashboard.

write_memory writes

Write per-user memory for channel sessions.

list_pipelines read-only

Introspect agent-defined long-running pipelines.

inspect_pipeline read-only

Inspect pipeline.

make_phone_call destructive

Place an opt-in outbound phone call.

Use cases

Deploy autonomous AI agents on your own infrastructure with governance and audit.

Manage a fleet of agents including scheduling, delegation, and monitoring.

Run agents written with Claude Code in production.

Coordinate multi-agent systems from a single YAML manifest.

Supported clients

Claude Desktop

Claude Code

Listed from the project's documentation, not tested by this site.

Overview

Trinity is a sovereign AI agents platform that lets you build and run proprietary systems, from fully autonomous businesses to self-improving cognitive systems, all governed, auditable, and on infrastructure you control. Each agent runs in its own isolated Docker container with real-time observability, fleet-wide scheduling, agent-to-agent delegation, and a tamper-evident audit trail. It supports multiple runtimes: Claude Code (Anthropic), OpenAI Codex, and Gemini CLI (Google). The platform includes a Vue.js frontend, FastAPI backend, and an MCP server (port 8080) exposing 90+ tools for external orchestration.

Similar servers

Context7 80 · B

Upstash's official server providing up-to-date third-party library docs for AI coding assistants

★ 62.9k · Tools 2 Compare with this →

NPM Sentinel MCP Server 79 · B

AI-powered NPM package analysis MCP server

★ 18 · Tools 19 Compare with this →

MartinLoop MCP Server 79 · B

An independent control layer that gives AI coding agents budgets, brakes, and receipts

★ 409 Compare with this →

GitHub MCP Server 76 · B

GitHub's official MCP server for managing repos, issues, PRs, and workflows via natural language

★ 33.5k · Tools 15 Compare with this →

Compare with other MCP servers

Source revision 43576279b43d Data synced 2026-10-11 Read the FMRS scoring method