Architecture

Built like an operations platform, not a chatbot with root access.

FrontBrain AI has a purpose-built orchestration layer, controlled tool execution, isolated workers, Remote Gateway support for private networks, and verification at every stage.

No mystery box. No blind automation. No “trust me bro” execution.

Layers

Three planes. One execution loop.

CONTROL

Control plane

The orchestrator drives the autonomous loop: intent classification, planning, workspace context, mutation gates, stall detection, verification, and audit logging.

EXECUTION

Execution plane

Toolrunner workers execute bounded tools. Stateful operations stay session-pinned where needed; stateless work can scale horizontally.

EDGE

Edge plane

The Remote Gateway runs inside customer networks and executes approved tools against private systems through an outbound-only connection.

Inside the loop

What happens between your prompt and the result.

Every task passes through a controlled lifecycle. The model does not simply improvise against production.

Intent classification

Classifies risk, cost, and operational impact before tools are called.

Workspace context injection

Adds active connectors, recent outcomes, open approvals, task state, and relevant memory to the model context.

Tool selection

Selects from bounded, approved tools rather than unrestricted system access.

Mutation gate

Classifies actions by what they do. Production-impacting steps require approval.

Read-before-write

Remote writes require a successful read or inspection of the same target first.

Shell hygiene

Commands are normalised for non-interactive execution where applicable.

Stall & platform-issue detection

Repeated same-error loops, no-state-change patterns, and gate loops pause the run.

Artifact fallback

When native artifact generation fails, fallback generation can use sandboxed scripts for documents, spreadsheets, PDFs, and slides.

Verification & final output

The final answer includes a Summary and Verification list with pass/warn/fail markers and evidence pointers.

Model routing

Front brain, escalation, BYOK.

FrontBrain AI uses a routing layer instead of relying on one model for everything. Routine execution can use efficient models. Hard reasoning can escalate to advanced models. Customers can bring their own OpenAI, Anthropic, or Google keys where supported.

The model is not the product. The execution system around it is.

router · policy · v3
Classify → Route → Escalate → Fallback
intent = infra_diag / risk = lowclassify
route → primary (Gemini 3.5 Flash)route
escalate on plan complexity ≥ 7watch
fallback → GPT-5.5 on provider outagestandby
BYOK override → customer keysoptional
Storage & state

Stateful where it matters. Isolated where it counts.

DB

Postgres and Redis

Conversation history, task state, approvals, and audit logs live in Postgres. Session events and runtime coordination use Redis.

Object storage

Generated files, screenshots, evidence, and artifacts are stored in managed object storage or customer-owned storage for private deployments.

Vault

Connector credentials live in Vault, scoped by tenant, user, workspace, and target.

Remote Gateway

Private-network execution without inbound ports.

The gateway is installed on a VM you control. It opens one outbound TLS WebSocket to FrontBrain AI and executes whitelisted tools inside your network.

Outbound only

No inbound firewall changes. No exposed management interfaces.

Whitelisted tools

Each gateway can only execute the tools explicitly enabled for it.

Audit at both ends

The cloud orchestrator and gateway both log actions so the records can be reconciled.

See deployment options
Ready?

See the loop in action.

Give FrontBrain AI a real task. Watch it plan, execute, verify, and record — with all the brakes on.

Less dumb. More done.