Let AI agents do real work. Keep humans, policy, and evidence in control.
Portarium is an open-source control plane for AI agents. Agents propose tool actions. Portarium checks policy, asks for human approval when needed, executes through controlled adapters, and records evidence so every decision can be tested, reviewed, and audited.
The practical goal is simple: you should be able to leave an agent working, know what it is allowed to do, and review the decisions that need you from a phone or web app without digging through logs.
Most agent demos let the agent call tools directly. That is fine for a prototype, but production teams need clear answers:
- Was this action allowed by policy?
- Did it need human approval?
- Who approved or denied it?
- What evidence existed at the time?
- What actually executed?
- Can we prove the safe, denied, retry, and isolation paths still work?
Portarium is built around that governance loop.
Agent proposes action
|
v
Portarium checks policy
|
+--> allow and execute
+--> ask for human approval
+--> deny or block
|
v
Evidence, result, and timeline are recorded
The core product is successful when automated tests prove that agents can route actions through this loop and Portarium enforces the expected behavior.
- Policy gates for agents:
Auto,Assisted,Human-approve, andManual-only - Human approval flow: rich context before risky actions run
- Mobile-friendly review path: approve or reject agent decisions from the Cockpit reference UI; hardening and polish continue
- Evidence trail: proposed action, decision, execution result, and audit record stay linked
- Controlled execution: adapters and action runners call external systems instead of agents using unchecked credentials
- SDK and API: agents, plugins, and tools can integrate through stable interfaces
- Reference Cockpit UI: operator screens for approvals, policies, runs, and evidence
- Repeatable tests: scenarios cover allowed, approval-required, denied, blocked, retry, and isolation paths
Portarium sits between agents and the systems they want to touch:
- An agent proposes an action through the SDK, plugin, or HTTP API.
- Portarium classifies the action using policy, risk, workspace, identity, and capability context.
- Safe actions can run. Risky actions wait for approval. Blocked actions do not run.
- Approved actions execute through a controlled runner or adapter.
- The decision, evidence, and result are recorded for review and testing.
These examples demonstrate the core governance loop. They are not required components of Portarium, and the business showcase ideas behind them are tracked as future work unless promoted into release specs.
More demo commands and media live in docs/how-to/run-cockpit-demos-locally.md.
Prerequisites:
- Node.js
^20.19.0or>=22.12.0 - Docker and Docker Compose
- Git
git clone https://github.com/45ck/Portarium.git
cd Portarium
npm ci
npm run dev:all
npm run dev:seed
npm run seed:cockpit-live:validate
curl -s http://localhost:8080/healthzdev:seed creates the live local Cockpit workspace ws-local-dev with pending approval apr-live-001, run run-live-001, linked work item wi-live-001, and evidence records for live UI walkthroughs.
Run the quickest governed-flow check:
npm run smoke:governed-runFor the guided path, use Hello Portarium.
Start here: docs/index.md
If you only read one page next: docs/index.md
Evaluate in 15-30 min: docs/getting-started/hello-portarium.md
| Need | Go to |
|---|---|
| Understand the product | Project overview |
| Know what is core vs future | Project scope |
| Run locally | Local development |
| Test the governance loop | Hello governed workflow |
| Run the core eval | Core governance eval |
| Trace evidence | Evidence trace tutorial |
| Understand the architecture | Architecture |
| Integrate an agent or system | Agent traffic controller |
| Use the HTTP API | HTTP API reference |
| See changes | Changelog |
| Contribute | CONTRIBUTING.md |
| Get support | SUPPORT.md |
| Report security issues | SECURITY.md |
Portarium is early and actively built. The core control-plane foundations are present: domain contracts, approval flows, policy tiers, evidence, OpenClaw/plugin integration, SDK surfaces, Cockpit reference UI, tests, migrations, and production scaffolding.
Remaining core work is mostly production hardening, security closeout, release hygiene, mobile Cockpit approval polish, live/eval proof, and keeping the tested agent-governance loop reliable. Larger demos, Growth Studio, mission-control UI, and pilot research are future work, not core blockers.
The @45ck/prompt-language verification-first runtime is a sister project Portarium is being designed to operate eventually. PL is in active experimental work (HA-HR1 hybrid local/frontier orchestration; see prompt-language/docs/strategy/program-status.md). The Portarium ↔ PL integration contract is not yet started by design — the trigger is portfolio-level positive evidence on PL's hybrid-efficiency tracker. Surface mapping and contract gaps are documented in docs/integration/prompt-language-runtime.md.
See project scope and roadmap.
Portarium is MIT-licensed open source. Contributions are welcome when they make the tested agent-governance loop safer, clearer, easier to run, or easier to verify.
Read CONTRIBUTING.md before opening a PR.
- Questions and ideas: GitHub Discussions
- Bugs and feature requests: GitHub Issues
- Security reports: GitHub Security Advisories
- Support guide: SUPPORT.md
Released under the MIT License.