v0.1.0 — open source, MIT licensed

Human-governed control plane
for AI agents

Let your agents run. Keep yourself in the loop. Approvals, constraint enforcement, database governance, earned trust scoring, and a full audit trail — in one open-source control plane.

Built by Project Black Box LLC · llms.txt · Disclaimer

OperatorBoard — Approval Queue
Approval Queue3
Agents4
Tasks
Org Chart
Analytics
AGENTS
ResearchAgentapproval_required · 94% approved
DBWriterscoped_autonomy · 87% approved
Approval Queue
3 actions pending review across 2 tasks
⚡ Webhook sent
db.schema.alter
Drop column "legacy_token" from users table
DBWriter · Task #47 · Backup: rds-snapshot-2025-04-28 ✓
CRITICAL
⚠ Destructive DB action — estimated 2.1M rows affected. Backup attestation verified 4 min ago.
http.get
Fetch research data from external API
ResearchAgent · Task #49
MEDIUM
file.read
Read config from /etc/app/settings.json
ResearchAgent · Task #49
LOW

Execution modes

Every agent starts conservative.
Autonomy is earned.

Assign each agent an execution mode. As it builds a track record — high approval rate, zero constraint violations — OperatorBoard suggests a promotion. You click the button.

observeAgent analyzes and reports. No actions taken.Read-only
proposeEvery proposed action goes to the approval queue before anything runs.Review all
approval_requiredTask-level approval gate. Default for every new agent.Default
scoped_autonomyAgent acts within an explicit constraint envelope. Violations are blocked.Earned

The problem

Most frameworks are fire-and-forget.

OperatorBoard adds the layer that's been missing.

Without OperatorBoardWith OperatorBoard
Agent acts immediately, no reviewAgent proposes; you approve before anything runs
No record of what ran or whyFull audit trail on every action, decision, and violation
"The agent said a backup exists"Independent backup attestation required before destructive DB ops
Shell commands bypass DB policyShell DB access auto-classified as write_destructive and blocked
Network and file access on by defaultOff by default; explicitly granted per task per agent
High-performing agents stay restricted foreverTrust score promotes agents automatically when criteria are met

Features

Governance is the product.

Not bolt-on safety features. The approval model, constraint enforcement, and trust system are what OperatorBoard is.

Approval queue

Multi-action review. Tasks stay blocked until every proposed action is decided. Webhook alerts when approval is needed — no polling.

🏆

Earned trust

Approval rate and violation tracking per agent. At ≥90% approval rate, ≥5 tasks, zero violations, OperatorBoard suggests promotion.

🔒

Constraint enforcement

Per-task locks on network access, file read/write, shell, and database tier. Case-variant and substring bypass paths are closed.

🗄️

Database governance

Four-tier DB access. Destructive actions require independent backup attestation with HMAC-signed integration ingest and replay protection.

📋

Full audit trail

Every action, approval, constraint violation, heartbeat, and integration attempt logged. Honeypot routes log scanner probes.

⏱️

Scheduled tasks

Queue tasks for future execution with ISO 8601 scheduling. Review results when you come back. Chain tasks with pipeline triggers.

💰

Cost analytics

Spend by day, task outcomes, and approval rates across your fleet. Per-agent budget hard stops included.

🌐

Org chart

Model reporting relationships between agents. Visualized as a collapsible tree. Kill switch suspends any agent immediately.

🛡️

Security hardening

SSRF prevention, timing-safe auth, integration replay protection, Caddy config with rate limiting and CSP — all included.

⚠️ Database governance: the hard problem

An agent with write access can silently wipe your database in seconds. OperatorBoard requires an independently-held backup attestation — not an agent claim — before any destructive database action can even be approved. Backup attestations are posted via HMAC-signed integration endpoints with provider binding, timestamp freshness enforcement, and replay protection. Shell-based DB access is auto-classified as write_destructive and blocked regardless of shell permission settings.


Earned trust

Autonomy as a reward,
not a starting point.

Agents don't start trusted. They earn it. You always make the final call.

01 — REGISTER

Start conservative

New agents default to approval_required. Every task needs operator sign-off before any action runs.

02 — OPERATE

Build a track record

OperatorBoard tracks approval rate, constraint violations, and tasks completed per agent across every run.

03 — EARN

Promotion suggested

At ≥90% approval rate, ≥5 tasks, zero violations — OperatorBoard surfaces a promotion suggestion. You click the button.


Quick start

Running in under five minutes.

One-click cloud deploy, Docker full stack, or local dev — pick your path.

npx operatorboard init

One-command setup — generates your API key, writes docker-compose config, starts the stack, opens the dashboard.

🚂
Deploy on Railway
API live in ~2 min
🎨
Deploy on Render
API key auto-generated
bash
# Zero-install quickstart — Docker required npx operatorboard init # Generates API key, writes docker-compose config, # starts the stack, and opens the dashboard. # Run with demo seed data: npx operatorboard init --demo # Dashboard → http://localhost:3000 # API → http://localhost:4100
View on GitHub →Read the docsProduction checklist

⚠️ Read before deploying

OperatorBoard is provided "as is" under the MIT License. Project Black Box LLC is not responsible for any actions taken by AI agents connected to OperatorBoard — including data loss, unauthorized access, financial charges, database changes, or any other harm — whether or not those actions were approved through the platform.

OperatorBoard reduces risk. It does not eliminate it. You are responsible for securing your deployment, understanding what each registered agent can do, and applying appropriate human oversight for your context and risk tolerance. Do not connect agents with access to production databases or sensitive systems without independent security review of your complete stack.

Read the full disclaimer → · Security policy →

Open source. MIT licensed.
Deploy it today.

Your agents are only as trustworthy as the controls around them.

View on GitHubDocumentationChangelog