A Go Binary That Manufactures Sandboxed AI Agents

What it does

FastClaw is an “Agent Factory” written in Go that creates and runs AI agents with distinct personalities, memory files, and skillsets. It handles the full operational stack—LLM communication, sandboxed tool execution, session management, and per-agent configuration—behind a web dashboard and an OpenAI-compatible API. The entire runtime ships as a single binary that uses SQLite by default and can be repointed to Postgres for heavier deployments.

The interesting bit

Instead of sprawling config files, FastClaw bootstraps purely from environment variables and stores all runtime state—provider keys, model overrides, IM channel tokens, and agent identity files like SOUL.md—in a database. The CLI and dashboard write to the same store, and the gateway hot-reloads on changes via SIGHUP, so you can provision an agent from the terminal and immediately chat with it in the browser without restarting the server.

Key highlights

• Single binary with SQLite by default; Postgres support for multi-pod deployments.
• Per-agent overrides for models, skills, and IM channels (Telegram, Discord, Slack) with isolated sessions per platform.
• Sandboxed tool execution via Docker or E2B, automatically syncing files back to durable storage after each call.
• Built-in user provisioning, agent quotas, and forkFrom cloning to support multi-tenant “buy a bot” workflows.
• MCP server support and a JSON-RPC plugin system for extending the built-in toolset.

Caveats

• On Windows, the CLI cannot signal the gateway to hot-reload; you’ll need to restart the daemon manually after configuration changes.
• The README’s configuration reference is truncated, so the complete list of system-wide config keys is not fully visible in the sources.

Verdict

Developers who want to self-host a multi-agent platform with per-user isolation and chat-channel integrations should take a look. If you need a managed service or a language-specific SDK instead of a self-hosted Go runtime, look elsewhere.