Map a prompt, workflow, document, codebase, or model. AI proposes. You decide.
Free account · your graph stays on your device until you share it
Already have work? Import a Mermaid flowchart, map a codebase, or open a template. Lucidchart and Miro import are not yet supported.
FlowGraph is one product across both surfaces. Pick the shortest route into a living graph; the same review and provenance laws apply.
Install the local FlowGraph app, then run flowgraph. Your graphs stay local unless you choose to share.
python3 -m pip install flowgraphappA free account is required on first run. The local canvas is free; Pro unlocks live CLI execution.
Start in the browser and let AI draft cards and connections as proposals you can inspect.
Try in your browser →Import supported Mermaid flowchart syntax as editable cards and edges—not a flattened image.
Open the converter →Begin with a real graph for decisions, roadmaps, knowledge maps, or construction workflows.
Browse templates →Every object has a story.
The element story is not a construction trick. It is how FlowGraph sees the world. Take a kitchen you know, and look again.
Every pipe, valve and wire can answer: what am I, why do I exist, who takes care of me, what happens if I fail. With a receipt on every answer. A kitchen today. A tower, a hospital, a factory the same way.
Illustration with example facts. In the app, every label is a real fact carrying a real receipt from your own model and documents.
Which side of AI do you want to be on?
FlowGraph doesn't replace your judgment.
It multiplies it.
The same Monday. The same AI. The only difference is where the line sits, and who holds the pen. Everything below is how FlowGraph keeps it in your hand.
FlowGraph has more to say than one homepage should ask you to absorb. The deeper chapters are preserved as focused Stories—each with a link you can send to a colleague or share on social media.
Why walls of fluent text and black-box answers fail when the decision actually matters.
Read and share → 6 min · connected workFollow a real object across drawing, model, schedule, field issue, and approval without losing the thread.
Read and share → 7 min · MCP + agentsHow local-first knowledge, governed writes, and MCP let your own AI work without silently rewriting truth.
Read and share →Be honest. When did you last read an entire AI answer?
You skim. You copy. You hope. Every day, AI hands you walls of fluent text, and somewhere inside them live the decisions that actually matter. Buried. Unranked. Unverified.
Reading is not reviewing. A wall of text hides its weakest brick. A graph puts every claim where you can see it, one card at a time.
the same answer, twice
what's inside the box
Do you know what your AI actually did?
It planned. It searched. It chose one path and quietly discarded four others. Then it handed you a paragraph and called it a day.
In FlowGraph, every AI run leaves an ordered trace: what it read, what it tried, where it failed, what it assumed. You can open any step, replay it, or hand it to a different model. No silent steps. Ever.
the flight recorder for AI work
What if the graph wasn't a diagram? What if it was the intelligence itself?
Not a picture of the work. The work. Humans see everything AI is thinking. AI reads everything humans have ratified. The canvas is the contract between you.
Plans, claims, and dependencies arrive as ghost cards on a canvas. Nothing hides in a paragraph, and nothing writes itself into your knowledge.
Approve, reject, or lock, one atom at a time. A locked fact cannot be overwritten by any model. The pen stays in your hand.
Your agents query the graph and get facts with receipts: who approved it, what grounds it, and how fresh it is. Verified context in, better AI out.
Everything you know becomes a card.
A canvas is a graph. The graph is made of cards, and every kind of knowledge gets one: goals, actions, risks, decisions, people, even whole documents. The lines between them mean something too: depends, feeds, blocks, approves.
AI proposes cards. You verify them. The map of your work builds itself.
Every card opens up like a watch. Eight atoms, each one approvable, lockable, and traceable on its own. Click an atom and watch its story, or just sit back: the card will tell it itself.
One card, one claim. Small enough to verify in a single glance.
Proposed, verified, or locked. Amber to green to law.
Which model, which run, which steps. Stamped forever.
The human who signed it. Trust always has a name.
What grounds the claim, one click from the original page.
How sure, stated plainly. Facts and assumptions never mix.
Facts rot. Every atom carries a review date and a queue.
How many cards lean on this one. Know before it breaks.
Most tools treat a card as pixels. FlowGraph treats it as a unit of trust.
Can you follow one real thing through everything?
In construction, a single door lives in a drawing, a spec section, a submittal, an RFI, a schedule, and a punch list. Six documents, six silos, and the person on site gets to play detective. Every industry has its own version of this.
Smart cards end the hunt. Drop a document into FlowGraph and it becomes a card whose contents other cards can point at. The same real thing, one thread, end to end. Change one link and everything downstream feels it.
One door. Six documents. Zero hunting. When RFI #87 changes, the install card knows before your crew does.
What happened here, and why?
Every element on a jobsite has a history: the clash, the RFI, the answer that cited a spec page, the decision with a name on it, the change order that paid for it. Today that history is scattered across five systems and the inbox of someone who rolled off the project in March.
The Element Story keeps it. FlowGraph anchors every record and every decision to the element it touched. Click a duct. Know its whole life. Pick your role, and ask your question.
Pin a comment to an element. The RFI, the change order, and the decision live in other systems, under other logins.
Answers fast, from your files. It has never seen your model, and it cannot point at the element it is talking about.
Link records inside their own walls. Export flattens every link. Cancel the subscription and the connections are gone.
FlowGraph keeps the why. Who decided, what it touched, what it cost, anchored to the element itself, with a receipt on every fact. It runs local-first, works offline in the trailer, and the record is yours after every subscription ends.
When the claim comes, you are not reconstructing the story from six inboxes. You have the receipts.
Sources: Navigant Construction Forum, Impact & Control of RFIs, 2013 · FMI + PlanGrid, Construction Disconnected, 2018 · Arcadis, Global Construction Disputes Report, 2025. Figures are industry studies, not FlowGraph claims.
One graph. Seven ways to see the work.
A roadmap in slides. The same roadmap in a spreadsheet. Again in a kanban. Again in a doc. Four copies, four owners, four ways to drift out of date.
In FlowGraph you capture it once. The same cards, the same atoms, rearranged live into whatever shape the question needs. Watch them move: nothing is copied, nothing is exported, and each view remembers how you arranged it.
Executive altitude for the board. Kanban for the standup. Timeline for the deadline. Outline reads like a document. Logos for the tool map. One truth, wearing seven outfits.
Ready to be the human in the loop?
Proposals appear as dashed amber overlays. Nothing is written yet; the graph is untouched.
One keystroke per atom. Approved facts turn solid and get stamped with your name.
Who proposed it, who approved it, what grounds it, when it expires. Attached to the fact, forever.
"Wait. Isn't this just...?"
Fair question. You already have tools with boxes and arrows. Here is where they stop, and where FlowGraph begins.
Notes tools store. Automation runs. Whiteboards draw. FlowGraph governs.
And whose knowledge is it, anyway?
Yours. Full stop. FlowGraph is local-first: your graph lives in your browser, works offline, and leaves your machine only when you explicitly say so.
Bring your own AI keys, from any provider. Exports are scrubbed of anything you didn't choose to share. Your judgment is your edge; FlowGraph never asks you to hand it over to use AI.
access by permission only
The graph is not documentation. The graph is the control system.
When agents do real work, they need two things from you: verified context to read, and a governed gate to write through. FlowGraph is both, over plain MCP.
apply_graph_patch proposes; your review inbox decides. An agent cannot corrupt what it cannot directly write.
Agents filter to human-approved knowledge only, and see exactly how fresh every fact is.
Grant read, write, or execute at the workspace, project, canvas, or single-fact level. Most specific wins.
Your agents drive FlowGraph over MCP, using the Claude or ChatGPT subscription you already pay for. FlowGraph drives your tools the same way, through a local server that keeps credentials off the browser.
// any MCP-capable agent, one call query_graph({ canvas: "launch-q3-platform", min_trust: "human_approved" }) // facts arrive with receipts attached { fact: "checkout depends on auth-v2", status: "verified" // locked, approved_by: "priya", sources: ["adr-017", "arch-review.pdf"], review_due: "in 12 days", blast_radius: 3 }
The plan was never the goal. The outcome was.
In the Agent view, the graph you approved re-arranges into an org-chart of runnable steps: every action gets a play button, a contract, and a status that cannot lie. Dry-run, live, and human-approved are three different greens — and anything mocked says so.
Review each step · auto within limits · or full access to the goal — granted per run, on one consent card that names the scope, tools, and spend ceiling.
Bounded retries toward each action's machine-checkable test. An AI opinion alone can't turn anything solid green.
Every non-pass produces a Run Report — what it tried, where it stopped, why, when, and exactly what's needed — assembled from the trace, so it can't confabulate.
Pyramid: everyone answers, one blends. Funnel: each refines the last. Per-layer provider, tokens, and cost — estimated before, recorded after.
$ flowgraph "export revit to ifc + nwc" --auto --budget 2 ▸ Export IFC via aps · confidence ●●●○ ✓ LIVE pass · model.ifc · sha256 ✓ · $0.41 ▸ Export NWC blocked · needs: Windows + NWC exporter no cloud route exists — Run Report attached flow: 1 pass · 1 blocked · $0.41 · every step receipted
also from your own agent: claude & chatgpt drive flowgraph over mcp — with the subscription you already have
Point it at a repo. Read the map before the code.
Paste a GitHub URL — or pick a local folder — and FlowGraph turns the source into a living graph: modules become cards, imports and calls become edges, entry points get a “start here.” A new engineer can walk it in an afternoon. Public repos need no sign-in; the code is fetched right in your browser and never touches our servers.
paste https://github.com/owner/repo ▸ 128 modules → cards imports & calls → edges ▸ entry points marked start here ➊➋➌ ▸ nothing lands until you approve the changes for commit history, churn & ownership: run flowgraph map locally
open the app → Map a codebase · works with TypeScript, JavaScript & Python today
You saw both sides at the start.
If you scrolled this far, you already chose your door.
One rule: you pay only for the things you choose to use. No seats held hostage, no features moved behind a paywall after you depend on them. Full pricing →
flowgraph run with the same Pro entitlementone seat · cancel anytime · 30-day money-back
the rules we won't break — governance basics stay free · your keys stay yours, we never resell inference · locked features stay visible, never hidden · lapse and you keep every byte and every free capability
Open the canvas, type one goal, and watch a verified graph take shape. If you scrolled this far, you already know why it matters.