Why PRAE The agent is autonomous. The execution isn’t.
Enterprise AI agents can execute at machine speed. But they cannot reliably carry enterprise context, policy, and accountability with them. PRAE is the control plane for agent execution — grounding agent actions in organizational reality, steering behavior before execution, enforcing deterministic boundaries, and producing verifiable evidence of what actually happened.
PRAE steers what the agent decides, and gates what it runs. Every capability on this page is labelled for what ships today.
The enterprise agent problem
Agents move fast. Enterprise execution has constraints.
Prompts and instruction files tell an agent what it should do. On their own they cannot guarantee that what it executes stays within policy — and afterwards, an enterprise has to answer for every action, not every intention.
-
What was the agent trying to do?
The call it requested — tool and target — on every row, beside a hash of the prompt it was given. Recorded, not inferred.
Shipping -
What was it allowed to do?
The pack and version in force; a signed pack names the key that signed it.
Shipping -
What did PRAE allow, steer, hold, or deny?
The verdict on each call, with the rule that produced it. Steering is recorded where it is configured.
Shipping -
What actually executed?
Whether an allowed or released call ran or failed, and whether a tool result was withheld.
Shipping -
Why was the decision made?
The clause, its stated reason, and every rule that matched.
Shipping -
Can the organization prove it afterward?
A hash-chained, signed ledger; any row exports as a receipt checked without our systems.
Shipping -
Can validated knowledge be reused by future agents?
Not yet. Your packs persist and govern every next run; promoting what a run validated into future agents’ context is still to build.
On the roadmap
PRAE sits at the execution boundary — between what the agent decides and what runs — where each of these can be answered at the moment of the call.
Who it’s for
One control plane, three people who answer for it.
Engineering leaders & CTOs
- Deterministic runtime control Shipping
Explicit rules decide each tool call before it runs — allow, hold for a human, or deny — strongest match wins.
- Agentic trust at scale Shipping
One policy for every agent on a governed runtime, reviewed or not. Roll it out in shadow mode, read the verdicts, then enforce.
- Measurable execution Shipping
Every decision is a row: what was allowed, held, refused or steered, by agent and session — exportable as OCSF events.
VP Engineering & platform leaders
- Turn existing instructions into control In part
PRAE reads the AGENTS.md, CLAUDE.md, .cursorrules and .windsurfrules an agent is given and proposes the enforceable part as rules for review — never applied unreviewed.
Cursor’s .cursor/rules directory is not read yet.
- Make knowledge compound On the roadmap
Today your packs live in the repository, versioned and hot-reloaded, and govern every next run. Carrying what a run validated into future agents’ context is not built yet.
- Fit existing workflows Shipping
Packs are files you review like code. prae test defends them in CI with a non-zero exit; decisions reach your SIEM as OCSF events.
CISOs & security leaders
- Runtime containment Shipping
Stop unauthorized actions before they execute.
Each tool call is refused, held for a human, or allowed at the execution boundary, before it runs. No probabilistic model in the path: the same call against the same pack gets the same verdict, with the clause on the record.
In full for agents that call judge(). Cursor, Windsurf, Claude Desktop and GitHub Copilot in part: file access through PRAE’s MCP server. Slack AI and Microsoft 365 Copilot not yet.
- Blast radius control In part
Limit what an agent can reach and change.
Rules bound the tools, files, hosts, commands and budgets an agent may use, grouped into Guardians. The Command Center’s blast-radius view shows what an agent can reach, and what PRAE allowed, observed, held or refused on each path.
PRAE contains the calls it judges; it does not prevent every compromise. Path and host rules do not apply to shell commands.
- Verifiable evidence Shipping
Know what the agent actually did.
Every decision and its outcome sealed in a hash-chained ledger, signed with your Ed25519 key. Any row exports as a receipt a third party verifies independently.
For CISOs & security leaders
Control agentic risk without slowing engineering.
AI agents can access sensitive systems and take consequential actions at machine speed. PRAE gives security teams a runtime enforcement layer that controls agent execution, limits blast radius, and produces verifiable evidence of what happened.
Agent surface
- Cursor In part
- Claude Desktop In part
- GitHub Copilot In part
- MCP In part
- Custom agents · judge() Shipping
PRAE · runtime control
Evaluate
Each call, against the pack in force.
ShippingEnforce
Allow, hold or deny, before it runs.
ShippingSteer
Withhold a result, and tell the agent why.
In developmentEnterprise systems
Your tools, files, systems and data. Only what PRAE allowed reaches them.
Verifiable evidence
- Decisions
- Interventions
- Outcomes
- Audit evidence
Security policy is enforced where the agent actually attempts to act. The agent remains autonomous. The execution remains protected.
What your team can check
- A refused call never runs: in the red-team run, a refused export delivers 0 bytes to its endpoint.
- prae verify re-derives the hash chain; an Ed25519 signature covers each row’s hash.
- Changing a protected field on any row breaks verification.
- Signed receipts verify offline, against a key you already trust.
A receipt authenticates what PRAE recorded. It does not, on its own, prove a side effect in the world.
Compliance evidence
PRAE produces evidence that can support your existing security, AI governance, and compliance controls.
- SOC 2 Mapped
- NIST AI RMF Mapped in part
- ISO/IEC 42001 Mapped
- EU AI Act Mapped in part
Evidence that supports, not certification: PRAE does not make an organisation compliant, and an auditor decides whether the evidence satisfies a control.
What each mapping covers →The PRAE difference
Control the execution — not just the conversation.
Each of these has its place, and PRAE works beside them. What none of them holds is the decision itself: the agent’s context and the policy, applied to a call before it runs.
- Prompt and instruction management
- Tells the agent what to do. The model may or may not follow it.
- Runtime monitoring
- Sees processes, files and sockets — some of it prevents — without the agent’s context.
- Post-execution detection
- Finds the action after it has run.
- Model-based guardrails
- A probabilistic judgment on the conversation, not on the call.
- Red-team validation
- Tests scenarios before deployment, not each call.
Steer before execution
In developmentWhen a tool result carries instructions aimed at the agent, PRAE can withhold it and tell the agent why before its next step. An operator can answer a held call with ALLOW, DENY or STEER from the Command Center. A steer never relaxes policy.
In development: demonstrated in PRAE’s local Command Center, not yet in the SDK, over MCP or on any agent surface. Content screening takes a pluggable engine, named on every row; the engine bundled today is a deterministic demo adapter.
Gate what actually runs
ShippingEvery tool call judged before it executes — shell, files, network, MCP and your own tools — allowed, held for a human, or refused, with the rule that decided.
In full in Node / TypeScript tool loops; for Claude Code, Cursor and Windsurf, file access through PRAE’s MCP server.
Prove what happened
ShippingEvery decision sealed in a hash-chained ledger, signed with your key. Any row exports as a receipt anyone can check without the ledger or our systems.
A receipt authenticates what PRAE recorded. It does not, on its own, prove a side effect in the world.
How PRAE works
Model decides intent. PRAE decides what executes.
PRAE is in the path, before execution — not watching after the fact. A refused call never runs, and the decision is on the record either way.
Agent intent
The model chooses a tool call.
PRAE · execution boundary
PRAE context
The agent, session and pack in force, and a hash of the prompt it was given.
In partPolicy / constraints
Rules over tools, paths, hosts, commands, arguments and budgets.
ShippingPRAE decision
- Steer
- Allow
- Hold
- Deny
Steer: in part
Execution
Only an allowed call runs. A held call runs only after an approval for that exact action.
ShippingVerified evidence
The decision sealed, chained and signed — whatever the verdict.
ShippingCompound
Validated context carried into the next run.
On the roadmapNothing to the right of the boundary happens unless the decision allows it. PRAE does not replace the model or choose its actions; it decides which of them run.
Compounding
Every validated run can make the next run better grounded.
Through context and policy — never the model. PRAE does not train models or change their weights. Three steps of the loop run today; closing it is on the roadmap.
- Ingest
Instruction files read with their file, line and layer; the enforceable part proposed as rules you review.
In part - Control
Each call judged against the pack before it runs.
Shipping - Prove
Each decision sealed in the ledger, with the clause that made it.
Shipping - Compound
Verify what a run discovered, promote it, and hand it to the next agent as context.
On the roadmap
The next run starts from what the last one validated
Built for the agent ecosystem
Not limited to one model, IDE or vendor.
There is no model in the enforcement path, so the policy does not depend on whose model the agent runs. One pack applies wherever PRAE is mounted — and where it can be mounted today is labelled below.
Developer agents
In partClaude Code, Cursor and Windsurf: instruction ingest and file access through PRAE’s MCP server — their built-in tools are not intercepted yet.
Learn more →Custom & homegrown agents
ShippingCall judge() from prae-gate in your own Node or TypeScript tool loop. Python and other languages are not supported yet.
Learn more →MCP
MCP & MCP servers · In partMCP Gateway · In developmentAny MCP-capable agent can mount PRAE’s MCP server for gated file access. A gateway in front of every MCP tool call is in development.
Learn more →Deployed SaaS agents
In partServices in Node or TypeScript ask the gate before each tool they run. Managed cloud agent platforms are on the roadmap.
Learn more →CI/CD & engineering workflows
In partPolicy tested like code: prae test fails the build when a pack stops holding its invariants, and prae check rejects a pack that would not load. Governing agents that run inside your pipelines is not covered yet.
Integration
Control without replacing the stack.
PRAE mounts at the seam your agents already have. What you can install today is on the left; what we are building is on the right, and labelled as such.
Available today
-
Developer-first integration
npm i prae-gate, then call judge() before each tool your agent runs.
Shipping -
Node compatibility
prae-gate runs on Node 22.19 or later, or Node 24 and up.
Shipping -
Instruction-file compatibility
AGENTS.md, CLAUDE.md, .cursorrules and .windsurfrules, read and proposed as rules.
In part -
MCP-aware
prae mcp: gated file reads for any MCP-capable agent or IDE.
In part -
CI/CD
prae test and prae check exit non-zero, so a pack is defended in the pipeline.
Shipping -
Local and deployed
On a developer’s machine or in a deployed Node or TypeScript service, through the same judge().
Shipping -
Deterministic policy
Packs as JSON or YAML, testable, signable, and rolled out in shadow mode first.
Shipping -
Verifiable evidence
prae verify, signed receipts, OCSF export for your SIEM, and a GRC evidence export.
Shipping
Active roadmap
- MCP Gateway In development
One boundary in front of every MCP tool call, not only file access.
- Built-in tools in Claude Code, Cursor and Windsurf On the roadmap
Judging those agents’ own tools before they run, not only their file access through MCP.
- Cloud agent services On the roadmap
Managed agent platforms and cloud workflows.
- Compounding On the roadmap
Verifying what runs discover and carrying it into future agents’ context.
Architecture
One boundary, in the path. Everything around it stays yours.
PRAE does not replace the model, the agent framework, or the infrastructure they run on. It is mounted where a call becomes an action.
- Yours
Agent / harness
Your model, your agent framework, your harness. It decides what to attempt.
- PRAE
PRAE execution boundary
Policy + context + steering, applied to each call before it becomes an action.
- Yours
Tool / system execution
Your tools, systems and data. Only what the boundary allowed reaches them.
- Yours
Evidence / ledger
Written by PRAE for every decision, chained and signed with a key you hold. Yours to keep and to export.
Give agents autonomy without giving up control.
Run the real policy kernel in your browser, or tell us about the agents you need to govern.