Stage0
STAGE0 - Runtime authorization for AI agents

Before your agent can call a tool,let Stage0 approve the action.

Stage0 is a decision API for agent runtime control. Send the proposed goal, tool context, and side effects, receive GO / NO_GO / DEFER plus ALLOW / DENY / DEFER verdicts, and keep final execution inside your own system.

Stage0 never executes tasks for you. Final execution control always stays with your system.

Built for MCP servers, OpenAI Agents SDK flows, LangGraph pipelines, internal copilots, and any workflow that needs an explicit decision before execution.

Use Stage0 as the authorization layer between model intent and real-world action.
{
  "goal": "Send a payout summary to finance and trigger the payout workflow",
  "tools": ["payments_api", "slack", "internal_db"],
  "constraints": ["human approval required", "business-hours only"],
  "side_effects": ["money movement", "external notification"]
}
DECISION
ALLOW
DECISION
DENY
DECISION
DEFER

Approves actions, not prompts

Evaluate proposed goals, constraints, and success criteria before execution starts.

Controls tool use and side effects

Gate database writes, external requests, file changes, webhooks, and destructive actions.

Built for governed agents

Return deterministic ALLOW / DENY / DEFER responses with audit-friendly reasons.

A control layer between plan and execution

How Stage0 fits into an agent loop

Stage0 is not another model wrapper or prompt template. It is a runtime decision point in front of tools, workflows, and high-risk actions.

1. Submit the proposed action

Send goal, tools, constraints, side effects, and execution context to `/check`.

2. Evaluate policy and quota

Stage0 scores risk, validates inputs, applies policy, and checks plan limits.

3. Receive a decision

Your app gets a GO, NO_GO, or DEFER decision, the paired verdict, and structured response fields such as issues and risk signals.

4. Execute only if your system agrees

Stage0 stops at judgment. The final action still belongs to your own runtime.

Implementation focus

Agent Tool / MCP Authorization

The clearest wedge for this project is Agent Tool / MCP Authorization: the decision layer that sits between agent intent and tool execution.

Project

stage0-agent-runtime-guard

View project on GitHub

Skill

stage0-execution-guard-skill

View skill on GitHub

OpenAI Agents SDK

openai-agents-sdk-tool-gate

View on GitHub

LangGraph

langgraph-stage0-gate

View on GitHub

MCP Server

mcp-server-stage0-authorization

View on GitHub

Where this fits first

Deployments and migrations

Stage0 fit

Gate production changes before CI/CD, rollout, or schema tools can run.

production deployschange windowsrollback checks
View scenario

Payments and external messaging

Stage0 fit

Force approval or defer high-risk actions before money movement or outbound communication.

payoutsrefundscustomer notifications
View scenario

MCP and internal tools

Stage0 fit

Put authorization in front of private repos, internal APIs, and sensitive control-plane tools.

MCP serversinternal copilotsresource-scoped access
View scenario
MCP and tool-calling agentsWorkflow automation with high-risk actionsInternal copilots connected to private systemsAI SaaS products with costly or sensitive operationsApproval flows for deletes, payouts, deploys, and writesTeams that need logs, auditability, and decision reasons

Prompts influence output. Stage0 authorizes action.

Prompt engineering can shape behavior, but it does not create a reliable authorization boundary. Stage0 gives you a separate control plane before real-world side effects happen.

Decision proof

Decision proof buyers can inspect

These are representative decision records. They show what a blocked or deferred run looks like before any real side effect executes.

BLOCKED RUN

Production delete with the wrong role

A destructive action is proposed without a role that satisfies the required policy.

decision: NO_GO
verdict: DENY
issues: ["ROLE_NOT_AUTHORIZED", "DESTRUCTIVE_ACTION_REQUIRES_APPROVAL"]
request_id: "req_block_01"
policy_version: "[email protected]"
Inspect path

DEFERRED RUN

Payout webhook missing approval evidence

The action is not denied forever, but it cannot continue until approval context exists.

decision: DEFER
verdict: DEFER
issues: ["APPROVAL_REQUIRED", "APPROVAL_TIMESTAMP_MISSING"]
defer_questions: ["Who approved this payout?", "Is approval still valid?"]
Inspect path

TRACE CHAIN

What operators can prove later

The value is not only the verdict. It is the trace you carry into logs, audits, and incident review.

request_id -> req_block_01
policy_version -> [email protected]
environment -> production
approved_by -> [email protected]
final_runtime_action -> execution_stopped
Inspect path

This is the kind of proof buyers can evaluate: what was proposed, why it was blocked or deferred, and which fields explain the final enforcement decision.

Production checklist

A buyer should be able to confirm these four things before treating Stage0 as a production control layer.

Open production behavior docs

Fail closed

If Stage0 is unavailable or returns an unknown decision, the side effect must stop.

Trace every decision

Carry `request_id` and `policy_version` into your runtime logs, approvals, and incident review.

Require approval context

High-risk actions should include approval state, actor role, and target environment before execution.

Keep final enforcement server-side

Stage0 decides, but your own runtime still owns the final execution boundary.

Choose a plan based on operational risk

Price the system around the risk of unauthorized side effects, tool misuse, and wrong actions consuming quota.

Free

Active
$0/month

For local prototypes and early policy experiments.

  • Basic ALLOW / DENY / DEFER checks
  • Starter audit fields
  • Dashboard access
Start free

Starter

$9/month

For production pilots with real tool-use guardrails.

  • Higher monthly check quota
  • Deeper decision fields
  • Priority support for integrations
Choose Starter

Pro

$19/month

For teams running agent workflows with real operational risk.

  • Expanded quota and analytics
  • Cost estimate support
  • Best fit for governed production agents
Choose Pro

Prompts are guidance. Stage0 is runtime authorization.

If your agent can trigger a real-world effect, it needs a decision layer that exists outside the prompt.