Why we built CrabTalk
Agent runtimes ship memory leaks, CPU spins, and privacy regressions because they're built like prototypes. CrabTalk is built like infrastructure.
Agent runtimes manage long-running sessions, execute tools with real permissions, and talk to APIs that cost money. But most are built like weekend projects — fast growth, no discipline, regressions shipped to production.
We built CrabTalk because an agent runtime should be engineered like a database, not vibe-coded like a hackathon entry.
The state of the art
OpenClaw is the most popular open-source agent runtime. 1.4 GB of node_modules. 2–4 GB Docker image. After 13 hours of uptime: 69.9% CPU, 1.9 GB RSS, effectively stuck. Agent workers can OOM your host — no limits, no caps. Subagents silently fall back to cloud models when credentials don't propagate. Your data goes to OpenAI because of a config bug.
Install size (MB)
Hermes takes a different path — Python, six execution backends, 40+ built-in tools. Each backend has its own setup, failure modes, and security model. One integration pulls 70+ packages. The team is already concerned about dependency sprawl.
These aren't edge cases. They're the consequence of building without the discipline required for production infrastructure.
8 megabytes
Not plus Docker. Not plus a gateway. Not plus a database. A single binary, zero dependencies.
Moving parts required to deploy
The daemon is a long-running process — closer to PostgreSQL than to LangChain. Sessions survive client disconnects. Background agents run without a terminal. Config changes take effect immediately — no restart, no lost tasks, no downtime.
Everything outside the core is an extension — a separate process, a separate crash boundary, the same protocol whether first-party or third-party. Memory recalls context before every agent run. Gateways stream events to Telegram in real-time. Monitoring observes every tool call. None of them are in the daemon binary.
Delegation creates an isolated session — own history, own tools, own context window. The parent gets only the compact result. No context pollution between agents. No unbounded growth. Multiple agents genuinely concurrent.
Install size vs. time to first agent
Performance doesn't degrade. No memory leaks after 13 hours. No CPU spins from unused code. The runtime was designed for concurrency from the start.