Reverse-engineering Claude Code reveals Anthropica's undisclosed PaaS platform "Antspace": Built in Baku, self-hosted, full-stack ecosystem already taking shape

According to 1M AI News's monitoring, ArcBox Labs founder AprilNEA published a technical analysis article. By running standard Linux tools such as strace and go tool objdump within the Claude Code Web session, the internal architecture of the Claude Code runtime environment was dissected layer by layer, revealing a previously undisclosed Anthropic internal deployment platform called "Antspace".

Reverse analysis revealed that Claude Code Web runs in a Firecracker microVM with a configuration of 4 vCPUs, 16GB of memory, and 252GB of disk. PID 1 is a custom init process, and the process tree is minimal. The container's core binary, /usr/local/bin/environment-runner, is a 27MB Go executable file with debug symbols intact, maintaining the package structure and function names from the Anthropic private repository github.com/anthropics/anthropic/api-go.

Within the tunnel/actions/deploy/ package, two sets of deployment clients were found: the known VercelClient and the previously unseen AntspaceClient. The latter implements a complete three-stage deployment protocol (create deployment → upload tar.gz build artifact → stream NDJSON status updates), with a version string prefix of `staging-`. A search across the entire web yielded no references to "Antspace" on the Anthropic official website, blog, GitHub, or job pages.

Also disclosed was the internal codename 'Baku', the claude.ai web application builder, which uses a Vite + React + TypeScript template and automatically provides six Supabase MCP tools (on-demand database creation, migration management, Edge Function deployment, etc.). The default deployment target is Antspace rather than Vercel. The code also includes enterprise-level Bring Your Own Cloud (BYOC) environment support, allowing enterprises to run the environment-runner in their own infrastructure, with session orchestration by the Anthropic API.

The author's assessment is that this architecture constitutes a vertical integrated closed loop from model to runtime to hosting, competing with Vercel/Netlify, Replit, and Supabase.