Agents become useful when they can work with the tools and context people already use. They become dangerous when every connection is treated as a blanket permission: read everything, write anywhere, and hope the model interprets the request correctly.
FlowGraph takes a different approach. The graph is both context and control surface. Agents can query approved knowledge with its receipts attached. They can draft new cards, links, and updates. Mutations still cross the same governed patch boundary as a human edit, so the proposed change can be inspected before it becomes truth.
Bring the AI app you already use
FlowGraph exposes tools over the Model Context Protocol (MCP). A local Codex or Claude client can connect to those tools and use the subscription attached to that AI app for its reasoning. The user can ask in natural language: list my canvases, map a plan, query verified risks, or propose an update. FlowGraph supplies the structured operations and the governed data boundary.
That does not mean one subscription pays for every AI feature everywhere. There are three deliberately separate paths. ChatGPT Voice in the desktop app can coordinate a Codex task that uses FlowGraph MCP, consuming the user’s ChatGPT Voice and Codex allowances. Local FlowGraph Voice uses the user’s own OpenAI Platform API key stored on their machine. The Voice button on the hosted FlowGraph web app uses FlowGraph’s server-side Realtime API key and is included as a capped Pro service.
Receipts travel with the facts
A normal retrieval system returns text. A governed query can return the statement, trust state, sources, approving person, freshness, and blast radius together. An agent can filter for human-approved facts and avoid building a plan on proposals that nobody reviewed.
fact: checkout depends on auth-v2
status: verified · locked
approved_by: Priya
sources: adr-017 · architecture review
review_due: 12 days
blast_radius: 3 downstream atoms
The receipt also makes disagreement productive. If an agent detects a conflict, it can point to the exact fact and evidence rather than replacing the graph with a fresh summary. A person can compare the old ground with the new source and make a visible decision.
Sign-in pairs identity; it does not transfer your subscription
FlowGraph sign-in identifies the person, their plan, and any shared workspace permissions. Connecting FlowGraph to Codex or Claude is a separate local MCP setup. The AI app continues to authenticate and meter its own subscription. FlowGraph continues to enforce access to the selected local vault or connected workspace.
This separation is a feature. It prevents provider credentials from being casually moved into the browser, keeps local keys on the user’s machine, and makes the payer for each experience explicit. A hosted team can choose the convenience of FlowGraph’s capped Voice service. A local user can choose their own Platform account. A Codex or Claude user can keep the AI subscription they already prefer.
Autonomy should be granted per outcome
Good agent systems do not hide autonomy in a global toggle. They name the goal, tools, scope, spend ceiling, and proof required for the run. A dry run can show the exact plan before live execution. A blocked step can produce a report instead of pretending the outcome succeeded.
The destination is not an agent that needs no human. It is a system where one human decision can safely steer many agents because context is verified, proposals are visible, and every meaningful step leaves a trace.