You already pay for your AI. Stop paying twice to use it.
CodeMonster is a desktop AI app platform: your agent drives a real backend — Postgres tables, forms, workflows, roles — through approximately 100 named operations over MCP. Your model, your context, your account. No tokens. No credits. No meter. You bring the AI subscription you already pay for.
Live-tested with Claude Code over stdio. The other guides are written from the MCP spec and publish as live once we have run each one end to end; any MCP-compatible agent connects over the same authenticated surface. Desktop builds are being finalized for launch — request access and we send the installer link the day the build for your operating system ships.
Your agent, your desktop, our platform.
The desktop app is where you sign in and pick up your connection details. Local clients spawn the MCP server on your machine over stdio; network clients reach the MCP endpoint over HTTP+SSE with the bearer token the app shows. Either way the platform authorizes every call as you.
The desktop app sits beside both paths, not inside them: sign-in, the spawn command, the token.
The local server is a client of the platform, not a copy of it: your solutions, data and deployments live on managed cloud infrastructure we operate, built primarily on AWS. Not self-hostable, and not new — a backend that has carried enterprise workloads for twelve years.
Your client spawns the server
Claude Code and other local clients start the MCP server as a child process on your machine. The desktop app shows the exact command; the config block is a few lines.
Network clients bring a token
Clients that connect over the network get a URL and a bearer token from the desktop app. Every request carries the token. The only unauthenticated endpoint is /health.
One control surface either way
Same named operations, same authorization, same transactions. The transport changes how the bytes travel, not what the agent may do. See the control surface
Fast because it is governed, not in spite of it.
The agent never freelances in your backend. It assembles governed platform primitives — objects, forms, workflows, grids, screens — through an authenticated control surface of approximately 100 named operations, and every one of them is bounded by the six rules below.
Authenticated on every request
A bearer token is validated against the platform on every call. A rejected request gets a 401 and nothing else: no tool list, no schema, no server info. The desktop app issues the token for your signed-in session; token lifetime, rotation and admin-side revocation controls are being finalized for launch.
The AI acts with your permissions
Your token is forwarded on every platform call and the platform performs authorization. The MCP server never decides who can do what — the agent has your rights, never its own.
Maps, not JSON
Agents never see raw solution, form, workflow, grid or screenflow JSON. Reads return compact maps; writes are named for the change (add-field-to-form, set-object-permissions). One carve-out: a stripped single-field view.
Staged, previewed, then committed
begin-solution-edit stages. preview-solution-edit shows an entity-level diff. commit-solution-edit lands it as one save; rollback-solution-edit discards it before it lands. Leave it idle for 60 seconds and the transaction rolls itself back.
Explicit tenant context
list-tenants, then use-tenant, before any data write. Skip it and the write fails fast with Tenant context required. There is no auto-select.
A freeze flag per solution
Turn allowMcpEdits off and every AI write to that solution returns 403 — a human-only window for production-critical work. The flag does not travel between environments.
For the person who has to say yes: the IT and security brief · The architecture detail
A brand-new agent reads the manual first.
The MCP server ships its own domain knowledge in-band. get-domain-knowledge returns the platform's conventions — build order, object prefixes, groups first, tenant selection, publish to land — so an agent that has never seen the platform builds like one that has.
Every tool carries defaults and a description written for an agent, and every result comes back with a hint about what to do next. The agent does not guess the next operation; the last one told it.
When something is off, the agent can file report-feedback with its recent calls attached, and a human QA queue reviews it. Yes, the monster reads the manual.
Bring the one you already have.
Any MCP-compatible agent can connect. We publish a client as live only after we have run the setup end to end ourselves; until then the guide is provisional and says so.
Claude Code
Setup guide, live-tested: the config block, the first calls, what the errors mean.
Open the setup guideClaude Desktop
Guide drafted from the MCP spec. It publishes as live after a live test; the snippet is a starting point.
Open the setup guideCursor
Guide drafted from the MCP spec. It publishes as live after a live test; the snippet is a starting point.
Open the setup guideChatGPT
Guide drafted from the MCP spec. It publishes as live after a live test; the snippet is a starting point.
Open the setup guideAny MCP client
Generic HTTP+SSE setup with a bearer token. Per-client steps publish as live after we test that client; any MCP-compatible agent connects the same way.
Open the setup guideThe control surface
Every operation family, the transaction lifecycle, and what agents never see — approximately 100 named operations in all.
See the operationsBring-your-own-agent is not the same as bring-your-own-bill.
Hercules publishes an MCP setup page for ChatGPT, Claude, Claude Code, Codex or Cursor — a genuine strength, and the same bring-your-own-agent idea we hold. It also ships a business bundle with commerce and payments, and enterprise options such as SSO and SCIM.1
The difference is the meter. Every agent interaction still spends Hercules AI credits, including sessions driven from your own agent, and credit auto-billing can move you into the next tier so a running agent does not stall.23 Free is personal, non-commercial use only.4
Your AI subscription is your only AI cost. We never touch the tokens, so there is nothing to mark up. One organization license covers every builder, every app and every environment; production traffic is not metered either.
The catch, stated plainly: we do not publish a price — the license is quoted per organization. Platform-built apps run on managed cloud infrastructure we operate, built primarily on AWS, and the backend is not self-hostable.
Compare with Hercules · How the meter works · What you keep if you leave
Sources · last verified 2026-08-14
- Hercules pricing page — verified 2026-08-14
- Hercules docs: AI credits — verified 2026-08-14
- Hercules docs: plans — verified 2026-08-14
- Hercules terms of service (free-tier commercial use) — verified 2026-08-14
Hercules' MCP setup page — the source for the client list above — is cited in full on the Hercules comparison. This category reprices often: if a date above is older than a quarter, treat the claim as unverified and check the vendor page.
The ones we get asked about the AI.
What AI do I need?
Any subscription that can drive an MCP-capable agent. Claude Code is the path we test and document first; other MCP clients connect the same way. The desktop app connects your agent to the platform's control surface of named operations, and you are building.
Do you see my prompts?
Your prompts go to your AI provider under your account. What reaches our platform is the sequence of named operations your agent decides to run — create-object, add-field-to-form, commit-solution-edit — each authenticated as you. We record tool invocations for audit and quality; we are not a party to your conversation with the model.
Which models work best?
Any model good enough to run an agent that follows tool descriptions. The platform teaches your agent its conventions in-band through get-domain-knowledge and per-tool hints, so a capable model becomes productive without a human explaining the platform.
What if the AI does something dumb?
On platform apps it cannot introduce code at all — it can only call named, guarded operations, with your permissions, and a solution can be frozen against AI edits entirely for a human-only window. Edits are staged in a transaction: preview-solution-edit shows an entity-level diff before commit-solution-edit writes once, and an unattended transaction rolls itself back after sixty seconds. Custom frontends are the deliberate exception — there the agent writes ordinary Vite/React, and every deployment is versioned and rolls back per environment.
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.