Concept models · Architecture views
Let your agent draw the model.
Koncept is a modeling canvas with an MCP server built in. Point Claude at your workspace and it creates concepts, wires relations and lays out architecture views, live, on the same canvas your team already has open.
Free plan, no card. MCP included on every tier.
Agents
Your agent is a member of the workspace, not a script.
The MCP server is not an export hatch bolted on the side. It writes through the same live document the canvas does, so when your agent adds a concept, everyone with the model open sees it appear.
28 tools, one vocabulary
Create and search models, add nodes and relations, batch a whole diagram in one atomic call, lay it out, render it as a PNG the agent can actually look at, lint it for dangling edges and overlaps, open branches, post comments.
It signs its own work
Give a key a persona and the agent posts as itself, "Claude · reviewer", rather than hiding behind whoever minted the key. Connect interactively over OAuth, or mint an API key for CI. Either way roles are enforced on the server: a viewer's agent cannot write.
Payment was folded into Invoice. I split it out and wired Invoice "settled by" Payment on a branch. Want me to merge it?
Yes, good catch.
What you draw
Two kinds of model, one canvas.
Concept models
What your words mean and how they relate. Every concept carries a definition and the other names it goes by, so the argument about whether a "payload" is the same thing as an "artefact" happens once and stays settled.
Architecture views
What runs and what talks to what. Seven shapes for services, databases, queues, external systems and actors, translucent zones to group them, and layer controls when things overlap.
Rules, attached to the thing they govern
A constraint is a node with a name and a sentence, connected by dashed edges to whatever it applies to. Pair it with cardinality on a relation: cardinality says how many, a constraint says what must hold.
Real examples in the margin
Pin up to five concrete values beside a concept. An abstract box called "Electronic address" means little until it sits next to "Peppol participant id" and "SDI codice destinatario".
Teams
Change a shared model without a meeting.
Branches are the review
Propose changes on a branch. It is a normal, editable model, so the team reviews it on the canvas instead of squinting at a diff. Merge it or abandon it. No approval states, no thresholds.
Comments on the thing itself
Threads anchor to a model, a node or a single relation, so "this one is wrong" points at something. Agents post and resolve them too.
Mark it good, come back to it
Stamp a model when it reads the way you want. Restore that exact state later in one click, which makes letting an agent rework a diagram a decision you can undo.
Your data
Leaving is a button, not a support ticket.
A tool that holds your team's architecture and cannot hand it back is a tool nobody senior will commit to.
Export everything
One JSON file with every model in the workspace, archived ones included, in the same format a single model exports in. It sits above the delete button on purpose.
Deletion you can undo
Deleting a workspace cuts access at once and is recoverable for 30 days. Deleting your account removes your name and email; the models your team relies on stay where they are.
Backed up nightly
Production is snapshotted every night, and the restore path is the same one used to move the data here in the first place, so it is tested rather than assumed.
Pricing
Free while I work out how to fund it.
Koncept is free to use right now. Every account gets every feature, with no limit on models, people or agents.
That is a deliberate choice, not a permanent one. The free tiers this runs on will not carry heavy use for long, so pricing will arrive. Nothing you build will be held hostage when it does: your models export at any time, in full, in a format that re-imports.
If your company wants to use this more widely, or wants a say in where it goes, that is exactly the conversation worth having now.
Start with "draw my product".
Connect an agent, ask it for a first model, and correct it on the canvas.