Inside Claude Science: How Anthropic's Daemon-Driven Harness Works
A technical breakdown of Claude Science's daemon-mediated host RPC, kernel authority split, and agent-harness architecture.
Anthropic's Claude Science, released in beta on June 30, 2026, presents itself as a simple local app but actually launches a localhost web interface backed by a daemon that mediates every analysis. The real innovation isn't in the UI — it's in the authority split between kernels, which hold working state, and the daemon, which holds history and permissions.
Reverse-engineering the runtime reveals that a host SDK is injected into each kernel after its first execute() call, assembled by the daemon from base and gated code fragments tied to the active agent profile. Calls like host.llm() and host.mcp() aren't real imports — they're synchronous RPC requests serialized over a private kernel protocol, with a singleton registered in sys.modules to make 'import host' appear to work despite no actual package existing. The install itself is layered: the signed app is just a 111 MB bootstrap, while the full ~3.9 GB installation is roughly 95% separately provisioned Python, R, and MCP execution environments.
For developers building agent harnesses, Claude Science offers a concrete case study in combining model capability, harness policy, and application-owned execution into a single scientific pipeline: code compresses orchestration without collapsing capability boundaries, artifacts persist as versioned evidence independent of conversation history, and review runs outside the main agent's self-reflection loop, with the power to block task completion.