Design Partner Preview · Self-hosted · Provider-neutral

Give your AI
executive control.

Prefrontal Node is the governed AI control plane for applications that act—routing models, controlling tools and memory, requiring approval for risk, recording every decision, and testing behavior before release.

Explore the control plane Build a governed workflow

RouteGovernObserveProve

5: control-plane
superpowers
1: governed layer
across the stack
0: mandatory
provider keys

Prefrontal Control Center

runtime ready

LIVE EXECUTIONsupport.refund.review

risk · low

ROUTE

intent→policy→model→tool

mock / verifier.v148 ms

POLICY DECISION

Allow with verifier

Tool scope and cost ceiling satisfied.

TRACE

7 spanshash linked

EVIDENCE

Release confidence

00:00.000intent.classifiedsupport · low risk

00:00.017policy.evaluatedverifier required

00:00.048execution.completedevidence captured

⌂Self-hosted controlYour infrastructure. Your provider keys.

◇Provider optionalityStart without keys. Add capabilities as needed.

◎Policy-first operationRuntime decisions remain governed.

⌁Evidence by designTraces, evaluations, and release proof.

THE CONTROL PROBLEM

Your AI stack can act.
Can it explain itself?

Adding a model call is easy. Operating an AI system that can choose providers, retrieve memory, invoke tools, spend money, and change real systems is a different problem.

“The gap is not intelligence. The gap is controlled, explainable execution.”

Prefrontal Node product thesis

Behavior is scattered

Prompts, policies, provider rules, retries, and tools drift across application code.

Actions outrun governance

Agents can reach systems of record before risk, approval, and cost controls catch up.

Failures become incidents

Provider outages, poisoned memory, unsafe tools, and policy conflicts are discovered in production.

Evidence arrives too late

Teams cannot reconstruct why an action happened or convert failure into a regression test.

THE FIVE SUPERPOWERS

One control plane.
Every consequential AI decision.

Prefrontal Node turns fragmented orchestration into an operating layer that a host application can inspect, govern, test, optimize, and trust.

Visibility

Every run becomes a replayable, secret-safe trace across intent, model selection, tools, memory, policy, cost, latency, verification, and final action.

Flight recorder
Replay and differential runs
OpenTelemetry export

Governance

Declarative policy controls providers, tools, actions, memory, cost, risk, and domains—with approval paths for consequential work.

Policy VM and simulator
Tool sandbox
Human approval gates

Simulation

Attack and test AI behavior against prompt injection, poisoned memory, tool misuse, provider outages, and policy conflicts before release.

Eval and red-team forge
Continuous assurance
Deploy verdicts

Optimization

Route with awareness of quality, cost, latency, provider health, risk, and domain context—while preserving deterministic fallbacks.

Outcome-aware routing
Budget-aware escalation
Provider health windows

Safe autonomy

Generate reviewable improvement proposals, policy suggestions, regression tests, and task bundles without silent production mutation.

Human review by default
Attested task packets
Bounded improvement loops

THE RESULT

AI that can act—
without becoming a black box.

See the Design Partner path →

ARCHITECTURE

A governed chokepoint
between intent and action.

Prefrontal Node sits beside or inside the host application. It does not replace business logic—it gives every AI path a common operating contract.

HOST APPLICATIONS

Prefrontal Node governed AI control plane

CAPABILITY LAYER

ROUTE

Choose a path that matches the task—not a vendor default.

Filter by capability, policy, provider health, cost, latency, context size, modality, risk, and execution profile. Fall back safely when a provider is unavailable.

single providerfallbackcascadeensembleplanner / workerverifier

CONTROL CENTER

See the brain.
Operate with context.

A branded, read-only-first operator surface makes runtime state, providers, execution history, audit events, policies, approvals, diagnostics, and release evidence visible without exposing arbitrary filesystem access.

✓ Viewer and operator role separation
✓ Tamper-evident event-chain verification
✓ Curated evidence and readiness surfaces
✓ No JavaScript dependency in the embedded runtime UI

Read the security posture →

CONTROL PLANE SIGNALlive

runtimeready

event integrityverified

policy packproduction.v3

provider modeoptional keys

node › inspect trace_01JZ9N7

intent       customer.refund.review
risk         medium
route        cheap-first → verifier
policy       approval.required
tool         crm.refund.preview
memory       2 governed records
audit        sha256 chain verified

VERDICT      conditionally safe
next          await operator approval

WHERE IT FITS

Built for applications
where AI has consequences.

AI-NATIVE SAAS

Add governed agents without rebuilding your product.

Embed the control plane through the Python runtime, HTTP API, or sidecar pattern. Keep product logic in the host app while Prefrontal owns routing, policy, memory, tool controls, traces, and release evidence.

Multi-provider model access
App-scoped policies and actions
Trace-to-evaluation workflows
Customer-controlled deployment

DEPLOYMENT

Meet the application
where it already runs.

Start local, embed in-process, run as a sidecar, or operate a dedicated service. Provider keys remain optional capabilities rather than startup dependencies.

⌘

In-process runtime

Use the Python runtime directly inside an existing service for low-friction integration.

from prefrontal import PrefrontalRuntime

▣

Container sidecar

Keep the node beside the host app with a clear network and state boundary.

prefrontal-node serve

↔

Shared control service

Expose governed capabilities to multiple applications through the FastAPI surface.

POST /execute · GET /traces

DESIGN PARTNER PROGRAM

Turn one consequential workflow
into a governed reference deployment.

TAHAI works with a small number of teams to integrate Prefrontal Node around a real application workflow, establish policy and approval controls, instrument trace and evidence, and leave behind a production-readiness plan.

Now accepting scoped design-partner conversations

30–45 DAY ENGAGEMENTOne workflow. Fully governed.

01
Architecture and risk discoveryMap models, tools, memory, data, actions, and failure boundaries.
02
Control-plane integrationConnect one host application through the runtime, API, or sidecar.
03
Policy and approval designDefine allowed routes, tool scopes, limits, and human gates.
04
Trace, simulation, and evidenceReplay behavior, attack the workflow, and capture readiness proof.
05
Handoff and next-stage roadmapDeliver an operator package, known limits, and commercial deployment plan.

Request an architecture session →

EVIDENCE, NOT THEATER

A release is not “ready”
because a demo worked.

Prefrontal Node is built around bounded passes, explicit gates, reproducible artifacts, checksums, attestations, and honest distinctions between what was implemented, tested, and externally proven.

BUILDClean package and install smoke01

TESTSuccess, failure, abuse, and regression paths02

PROVETrace, policy, audit, scan, and restore evidence03

ATTESTExact claims, limits, checksums, and known gaps04

FAQ

Clear answers.
No black-box language.

Is Prefrontal Node another model gateway?

No. Routing is one layer. The product combines provider orchestration with policy enforcement, governed memory, tool approval, traces, simulation, evaluation, optimization, and release evidence.

Does it require a specific model provider?

No. Provider keys are optional capabilities. The node can start and demonstrate core behavior with mock providers, then add commercial or open-model providers as required.

Can it run in our infrastructure?

Yes. The intended commercial path is self-hosted and customer-controlled, using an in-process runtime, container sidecar, or dedicated service pattern.

Does it replace our application logic?

No. Your host application retains business logic and systems of record. Prefrontal supplies the governed operating layer around AI decisions and actions.

Is it generally available?

The current commercial posture is Design Partner Preview. Broader GA claims are gated on external CI, deployment, scan, restore, provider, security, and customer evidence.

NODE.TAHAI.NET

Your application has AI.
Give it a prefrontal cortex.

Bring one workflow, one set of risks, and one desired outcome. We will map the control plane around it.

Give your AIexecutive control.

Your AI stack can act.Can it explain itself?