Manta Make
An AI coding environment that runs the real Claude Code agent on self-hosted models — describe it, watch the agent build it, publish it
Manta Make is an internal technology demo. Because it executes arbitrary code, access is limited to the Careful Dream team.
About Manta Make
Manta Make is a web-based AI coding environment. You get a chat, a code editor, and a live preview — and behind the chat runs the real Claude Code agent, editing a real project on a persistent volume and streaming every step of its work back to the browser.
v2 is a ground-up rethink of the engine: instead of a hand-rolled language-model loop calling a hosted API, it bakes the Claude Code CLI into the image and drives it with self-hosted open-weights models on our own GPUs. The result is a genuinely agentic development tool that costs nothing per token and never sends your code anywhere but the cluster.
The agent writes a Pomodoro timer — code editor, chat, and live preview side by side.
A tap-for-sparkles app built end-to-end with v2 — generated and iterated entirely through chat with the self-hosted agent, then published to a public URL.
What's New in v2
A real agent, not a chat-to-code loop
v1 piped a raw LLM through a hand-rolled loop. v2 bakes in the Claude Code CLI itself — a full agentic coding tool that reads the project, plans, edits multiple files, runs tools, and iterates on its own until the change is right.
Self-hosted models, zero per-token cost
The agent is driven by an open-weights 27B model (Qwen) running on our own GPUs behind an in-cluster router. No LLM API fees, no subscriptions, and no code or conversation ever leaves the cluster.
Code and conversation that persist
The agent works on a real project tree on a persistent volume, not a sandboxed in-memory sketch. Source code, chat history, and full session context survive pod restarts — pick up exactly where you left off.
One-click publishing
Bundle any workspace into a static site and serve it at a public URL in one click. Share the finished app without sharing the workspace.
Passkeys, invites, and sharing
Sign-in is WebAuthn passkeys (no passwords, no API keys to protect). Access is invite-only, with collaborator sharing and unlisted links so you can hand a live workspace to a teammate.
Many sessions per workspace
Each chat is its own resumable agent session on the same project. Branch an experiment in one chat, keep the mainline in another, and reload either with its full history.
Key Features
Conversational Development
Chat with an agent that edits your actual project files — React, TypeScript, or vanilla — across as many files as a change requires.
Live Preview
The preview re-renders as the agent changes files, so you watch the application evolve in real time.
Real File System
File tree, tabbed editor, and syntax highlighting over the same source of truth the agent edits — inspect or tweak anything it writes.
Visible Agent Work
Every tool call, edit, and result streams into the chat as it happens, so you can see — and stop — exactly what the agent is doing.
Workspace Persistence
Files, chats, and published sites live on a persistent volume; a workspace outlives its pod and its reboots.
Publishing & Sharing
Publish to a stable public URL, share the workspace by link, or add collaborators by name.
How It Works
- 1Sign in with a passkey and open (or create) a workspace
- 2Describe what you want in plain English — the agent reads the project and starts editing files
- 3Watch its tool calls stream in while the code and preview update as files change
- 4Iterate by chatting — each session resumes with full context, so follow-ups are instant
- 5Publish with one click to share the running app at a public URL
Under the Hood
Agent Layer
The Claude Code CLI is baked into the container; each turn streams structured events (tool calls, edits, results) to the browser over SSE and resumes by session id.
Model Layer
Self-hosted Qwen served by an in-cluster router. The CLI's model aliases (Sonnet/Opus/Haiku) map onto our own GPU fleet, sized per model for context and speed.
Storage
SQLite for identity and workspace metadata; a persistent volume for the live source tree, agent transcripts, and published site bundles.
Delivery
Go + React (Sandpack) frontend, esbuild-based publish pipeline, deployed to Kubernetes with CI/CD from Git.
What We've Built With It
Rapid Prototypes
Turn a napkin idea into a clickable demo in minutes
UI Experiments
Test layouts, animations, and component designs interactively
Learning Tool
Explore patterns by asking the agent to explain and demonstrate
Interested in AI-Powered Development Tools?
Manta Make is an internal demo, but we build custom AI-integrated tools for teams. Let's talk about what we can create for you.
Contact SalesHave a project like this one?
We'd be glad to talk through your goals, your constraints, and whether we're the right team for the work.
