LUMENFLOW · PLATFORMMANUAL · GOVERN, APPROVE, EXECUTE, PROVE

DISPATCH · ONE SYSTEM

policy first.
approval at the boundary.

LumenFlow is one control plane for AI agent work: what’s allowed, where a human signs off, and what evidence is left behind — wherever the work actually runs.

Architecture

One system, three layers

Wherever the agent and wherever the compute, the same policy, approval, and evidence pipeline sits in the middle.

AI / Agents

SidekickClaude CodeCodexCursorMCP clients

LumenFlow

Policy

Approval

Evidence

Compute / Tools

Cloud modelsYour compute (BYOC)Connected tools

Policy flows down · Evidence flows up

GovernPolicy before execution.

Scope, spend limits, and quality gates are set before an agent touches anything. The control plane decides what's allowed — it isn't just told what happened afterward.

ApproveHumans stay in control when risk crosses a boundary.

Routine work keeps moving without friction. The moment an action crosses a policy boundary — spend, data access, a production change — it pauses for the person who owns that decision.

ExecuteRun through Sidekick, external agents, or connected compute.

Work happens wherever it makes sense: Sidekick's chat and playground, connected runtimes like Claude Code, Codex, Cursor, and MCP clients, or compute you connect yourself. The same policy and evidence layer wraps all of it.

ProveAn independently inspectable evidence trail.

Every governed action leaves a hash-chained, per-workspace sealed evidence record — what happened, why it was allowed, and who approved it — ready to export and inspect.

What operators inspect

Every important decision
should be legible.

Teams need more than a successful run. They need to know why a change was allowed, what evidence was attached, and whether a person was asked to approve it along the way.

Proof of execution

Policy resultAction allowed only after scope and budget checks passed

Approval pathHigh-risk step paused until the designated approver signed off

Evidence retainedReceipts, traces, and the final outcome attached to the task

Export readyJSON and CSV audit output available for operators and compliance

Connected AI

Agents you already run, governed the same way

Sidekick is LumenFlow’s own client. Everything below is someone else’s client, connected in as a governed runtime.

Claude Code

Connected runtime

Anthropic's coding agent, enrolled as a connected runtime that reports sessions, events, and evidence back to your workspace.

Codex

Connected runtime

OpenAI's coding agent connects the same way — through the control-plane SDK, under the same policy and evidence layer.

Cursor

Connected runtime

The Cursor IDE agent enrolls as a connected runtime, so its work is governed and evidenced like any other.

MCP clients

Open protocol

Any Model Context Protocol–compatible tool works inside LumenFlow governance without a custom integration.

ChatGPT

OpenAI Apps pilot path

A client start-point for approved delivery review — a pilot path today, not the governance owner.

See it work

Each tour starts with a real product surface.

Pick a motion, then compare the same task with and without LumenFlow.

Your AI just booked a $400 flight.
Who approved that?

Sidekick handles everyday tasks — travel, expenses, scheduling. Without governance, you have no idea what it accessed or spent.

Sidekick · Chat + Approvals

Real LumenFlow Sidekick chat surface showing the consumer-first entry banner and chat shell.

Real Sidekick chat surface with the consumer-first entry and the governed workspace shell.

Captured from the real app in a local seeded workspace

Why governance matters

Without LumenFlow

The task completes, but nobody can reconstruct what happened or why it was allowed.