Plan with your AI agent.
Without losing the plot.
Project tracking for teams building with AI agents. Claude Code breaks down your roadmap, drafts the specs, and fills the sprint board. Every change it makes shows up as a diff you read and approve before it counts. You get the speed of an agent without handing it write access to your project.
Plugs into Claude Code over MCP. Nothing lands without your sign-off.
From a rough idea to a sprint board you can run
Every project starts with a person. The agent does the filling-in; you make the calls that matter. Here's how a half-formed idea turns into tracked work.
You start it
You jot down the rough idea: a one-line vision, a half-baked spec, a roadmap item. The first move is always yours. The agent won't invent goals you never asked for.
The agent fills it in
Claude Code pulls only the context it needs over MCP, then drafts the specs, breaks the roadmap into chunks, and lays out the sprints and tickets.
You see what changed
Everything it wants to change shows up as a plain diff against what's there now, with the reasoning sitting next to it instead of buried in a log.
You approve it
Hit approve, or set rules that wave the small, safe changes through on their own.
It keeps itself clean
Throwaway docs expire on a timer; the canon worth keeping stays. Your specs don't slowly turn into a junk drawer.
Three guarantees, not features
You set the direction
Where the project is headed is always your call. The agent expands what you seed. It never decides the goal, and never invents one you didn't ask for.
Nothing lands without you
Every change is staged as a proposal you approve. Deletes and anything promoted to canon can never be auto-approved — they always wait for a person.
Context, not scattered files
The agent reads durable canon over MCP — not a heap of .md files nobody can find three weeks later. Working notes stay out of its view until you promote them, so it works off your best knowledge, not your half-thoughts.
One kind of doc, sorted by
how long it should live.
Under the hood, every entry in the knowledge base is one thing: a ContextDoc, tagged by kind. That tag decides whether a doc is permanent canon or temporary working notes, and whether the agent gets to read it at all.
- The important stuff sticks around. Specs, decisions, and conventions stay visible to the agent and only get flagged as stale once they're old enough to be worth a second look. Nothing critical quietly disappears.
- The throwaway stuff cleans itself up. Plans, research, and scratch notes all come with an expiry date and delete themselves, so you're not digging through last quarter's dead notes.
- Links run both ways. Open a spec and see its tickets. Open a ticket and jump back to the spec. Works the same in the app and over MCP, and every doc keeps its full history.
39 tools that match how work actually happens
The tools map to things you'd actually ask for, like breaking a roadmap item down or pulling the context for a ticket, rather than generic create-read-update-delete. Point Claude Code at Flowbord and it reads only what's relevant to the job, ranked so the most useful docs come first. Near-duplicates get caught before they ever reach your review queue.
Reads come back right away. Writes get staged as a proposal for you to look over. The one exception: throwaway working notes the agent saves on its own — and those expire on a timer.
See exactly what the agent did, and why
Every change the agent makes is recorded with its reasoning attached: what it did and why, in one place. That account lives next to the proposal, never inside your specs, so your documents stay lean instead of ballooning every time the agent touches them.
Grouped, diff-based review
Proposals arrive grouped by the job that created them, each shown as a plain diff against what's there now. You review a whole unit of agent work, with the goal it set out to hit and every change it made, instead of a pile of orphan edits.
Trust rules you control
You decide how much rope the agent gets. Small, reversible changes can go through on their own. Anything that deletes data or promotes to canon always stops for a person. The agent never gets autonomy you didn't hand it.
Every call audited
Every tool call the agent makes is recorded. Each change keeps its authorship — which agent and which key produced it — right next to the person who approved it, so you always know who proposed something and who signed off. Anything that looks like an API key is scrubbed from the log automatically.
Fail-loud tenant isolation
Each tenant is walled off at the database layer, and if a query ever looks like it might cross that line, it fails loudly instead of leaking. One API key per organization, with roles from Developer through Manager and Admin to Owner.
And the planning underneath — roadmaps, sprints, a kanban board, and burndown snapshots — all linked, so the agent's proposals land on a board you can actually run.
I built Flowbord because I needed it.
I'm Dennis. Twenty years as the person who owns the hard technical calls and still writes the code — architecting high-throughput data platforms, leading migrations, taking a product through its acquisition by Saint-Gobain. Alongside that I've built and run my own SaaS products solo, on infrastructure I run myself, with real users and real GDPR audits. These days I work with AI agents every day: custom MCP servers, self-hosted models, Claude Code in the loop. Flowbord is the planning tool I kept wishing I had — fast like an agent, on a leash I hold.
Like everything I ship, it's made and hosted in the EU, on EU-owned infrastructure with no US parent. And since your own Claude Code makes the model calls, your prompts and code never pass through me. You shouldn't have to choose between good tooling and keeping your data in Europe.
From a message bus handling 50,000+ devices to a handful of my own SaaS products — AppLogger, Shelfware, Styr På — that I build, run, and support end to end. Flowbord is the one I built to plan them all.
Questions, answered straight
What is Flowbord?
Project tracking for teams building with AI coding agents. A human seeds an idea; Claude Code, connected over the Model Context Protocol, expands it into specs, roadmap breakdowns, sprints, and tickets — as proposals you review and approve.
Does Flowbord work with Claude Code?
Yes. Flowbord runs an MCP server with 39 tools. Reads come back right away; writes are staged as proposals for your review. You connect it once with claude mcp add and point the agent at your project.
Can the agent change my data without my approval?
No. Every agent write is staged as a proposal. Deletes and promotions to canon can never be auto-approved and always wait for a person. You set trust rules to wave through the small, reversible changes you don't want to babysit — the only thing an agent saves on its own is throwaway working notes, which expire anyway.
Does Flowbord touch my code?
No. Flowbord plans the work — the agent proposes, you approve. Your code stays in your own repository.
How is this different from pointing Claude Code at a folder of .md files?
A folder of Markdown rots, and three weeks later nobody can find the file that matters. Flowbord serves structured context over MCP, ranked so the relevant docs come first, and shows every change as a diff before it lands. Working notes stay out of the agent's context until you promote them, scratch expires on its own, and specs link both ways to their tickets with full version history.
Is Flowbord free?
Yes — free while it's in beta, and no credit card is required to start.
How is my data isolated?
Your data is walled off at the database layer, not just in application code. If a query ever looks like it might cross between tenants, it fails loudly instead of leaking. Each organization gets its own API key, and people get roles from Developer through Manager and Admin to Owner.
Where is my data hosted?
In the EU, on EU-owned infrastructure with no US parent company in the ownership chain. An “EU region” run by a US-owned provider still falls under US legal reach like the CLOUD Act and FISA 702 — ownership, not the region label, is what keeps your data out of that reach. Flowbord is run by an EU-based founder and is GDPR-compliant by design.
Do my prompts or source code leave the EU?
No. The model calls run inside your own Claude Code, on your side — Flowbord never sees or forwards your prompts or your source. We serve structured context and store your project data (specs, tickets, proposals, audit records), and that stays in the EU. Anything that looks like an API key is stripped before it's ever written to the audit log.
Try it on your next idea
Point Claude Code at your roadmap and let it take the first pass. You sign off on everything before it counts. Free while we're in beta.
Create your workspaceNo credit card. A few minutes to connect Claude Code.