One desktop app. Your agent. A governed backend underneath.
CodeMonster is a desktop app that connects the AI subscription you already pay for to a real application platform over MCP. Your agent works through approximately 100 named, authenticated operations — never raw solution JSON, never freelancing. Builders get governed primitives on real Postgres tables, real environments, and an HTTPS URL.
No tokens. No credits. No meter. You bring the AI subscription you already pay for.
Desktop builds are being finalized for launch — the download page has the access list and the installer link.
Five layers between your prompt and a running app.
Your agent — Claude Code today, any MCP-compatible client — reads your prompt and decides what to do. It runs on your own AI account: your subscription or your API key, your rate limits, your bill.
The desktop app signs you in, connects that agent to the platform's MCP server, and shows the operation log as it happens. Your prompts go to your AI provider on your account. What reaches the platform is the sequence of named operations your agent decides to run.
The MCP server exposes approximately 100 named operations. Every request carries your bearer token and is validated on every call. Reads return compact structural maps; writes are named for the change they express — add-field-to-form, set-object-permissions. Never raw solution JSON.
Platform primitives — objects, forms, workflows, grids, screens — are assembled on a backend that has carried enterprise workloads for twelve years, with your records in real Postgres tables you can query. The rest of the stack is on the security architecture page.
Environments and the renderer put the result in front of your end users. Custom frontends move develop → staging → production, one gate at a time. Platform apps skip that chain: publish-solution copies the working copy to testing and production in one step, and the renderer serves them on web and mobile.
Platform apps, custom frontends, and one thing still on the roadmap.
Assembled from primitives through MCP
Your agent creates objects, forms, workflows, grids and screens through named operations. The platform renders the result for web and mobile. There is no generated backend code to review or maintain.
See the primitivesVite + React, shipped through the same chain
When the renderer is not enough, the agent writes an ordinary Vite + React + TypeScript app — or any framework that builds to static output — and ships it to develop, staging and production.
See custom frontendsServer-side functions
On the roadmap, not shipped. Status and scope live on the roadmap page.
Read the roadmapStaged. Previewable. Reversible.
Staged. begin-solution-edit opens a working copy of the solution. Every write after it lands in that copy, not in your app, and out-of-band writes are blocked while that one transaction is open. An abandoned transaction rolls itself back after sixty seconds of inactivity, so nothing half-finished sits open against a live solution — commit it or roll it back before you walk away.
Previewable. preview-solution-edit returns a human-readable, entity-level diff — “4 entities changed”, named — never a JSON patch. diff-against-stored runs a three-way drift check before you commit, in case someone else saved in the meantime.
Reversible. commit-solution-edit lands every staged change as a single save; rollback-solution-edit discards them all. publish-solution then copies the working copy to testing and production. Every step in the log above is a real operation name.
Humans and agents co-edit safely.
One writer at a time, many observers. That is the model, stated plainly — not multiplayer editing.
Same operations, one transaction per solution.
A person in the platform and an agent over MCP call the same named operations with the same permissions. While an agent's transaction is open, other writes to that solution are blocked and told so. A per-solution flag can freeze AI edits entirely for a human-only window: every agent write fails with a 403 until you lift it.
A single-writer edit session per app.
Writes to a working copy require the edit-session lock; a write without it gets a 423 and nothing changes. Reads are always free, so a person can watch an agent's edits land in the file tree as they happen. Takeover is explicit, and an idle session's lock expires on its own after two minutes without a heartbeat.
What every app ships with.
The parts of a business app that are the same every time are already built. Boring, on purpose.
- Authentication on every request — a bearer token validated on every call; a bad token gets a bare 401.
- SSO for your end users — plug in Auth0, SAML, Firebase or Google, so people sign in with the identity they already have.
- Groups and permissions — gate screens, actions and grids, evaluated server-side.
- Tenant isolation — row-level, on every object table, injected into every query.
- A change log you switch on per object — once it is on, every row-level change is recorded for people and for
get-instance-changelog.
- An HTTPS URL for every app — custom frontends get a develop, staging and production address; platform apps are served over HTTPS by the web and mobile renderer.
- A web and mobile renderer for platform apps — with the field-service patterns that ship today: offline mode, location, photo and signature capture.
- One-click deploy — publish or promote from the app or from your agent.
- No DevOps required — no build pipeline to own, no servers to patch, no on-call rotation to staff.
Apps run on managed cloud infrastructure we operate, built primarily on AWS. The backend is not self-hostable. One organization license covers every builder you have, and nothing on this list is metered — the only bill that moves with usage is the AI subscription you already pay for, and it is not ours.
The control surface, the chain, the brief, the bill.
The control surface
Your model, your AI account, and the named operations the platform lets an agent run.
See the operationsDevelop, staging, production
One-way promotion, owner/editor/viewer, non-destructive rollback, an append-only log.
See the chainFor the person who has to say yes
Auth on every request, RBAC, tenant isolation, and a plain list of what we do not claim.
Read the IT briefOne organization license
Every builder included, nothing metered. You bring the AI subscription you already pay for.
See pricingWhat the agent builds, and where apps run.
Does the AI generate my app's code?
For platform apps, no — that is the point. The agent assembles your app from governed platform primitives (objects, forms, workflows, grids, screens) through named operations, so there is no pile of generated backend code to review and maintain. Where you want a custom frontend, the agent does write ordinary Vite/React code, and it ships through the same versioned pipeline.
What kinds of apps can I build?
The categories that ship on the platform today: field-service and inspection apps, checklist and compliance apps, CRM-style apps, workflow automation and multi-step approvals, data-entry and public intake forms, business processes with PDF generation and external system calls — plus custom React frontends and dashboards over the same data. See use cases and blueprints.
Do apps work on mobile?
Platform apps render in the platform's web and mobile renderer, and field-service patterns — offline mode, location tracking, photo and signature capture — are part of what ships today. There is no native app-store build: the renderer is how a platform app reaches a phone. Custom frontends are whatever you build them to be.
Can humans and agents work on the same app?
Yes, safely. Platform solutions are edited through the same named operations whether a person or an agent is calling them, with one open transaction per solution at a time. Custom-frontend working copies use a single-writer edit session, so an agent and a person never clobber each other's files.
What environments do I get?
Custom frontends promote through develop, staging, and production — a strict one-way chain with role-gated promotion and per-environment rollback. Platform apps publish from a working copy to testing and production snapshots, with frozen backup snapshots for branching. Every push is one click; there are no pipelines to write.
Where do my apps run?
On managed cloud infrastructure we operate, built primarily on AWS. You do not provision servers, write deploy pipelines, or hire DevOps — dev, staging, production: one click each. The backend is not self-hostable.
Bring us something you want built.
CodeMonster is invite only while we work directly with early adopters. Request an invite and we will meet, scope what you want to build, and get you productive. You bring the AI subscription you already pay for.