Dispatch a task to a shared machine running the Hermes Agent. Poll terminal status, fetch transcript evidence, decompose the claims, and verify the outcome — from any MCP-capable client.


Claude Code, Codex, Ollama, Pi — any MCP-capable client sends a task over the authenticated bridge. Hermes owns the execution.

Tasks run on a shared satellite machine running the Hermes Agent. Every client orbits the same execution layer — deterministic, auditable, real.
Poll for status until the run reaches a terminal state, then fetch the transcript. No callbacks, no guessing — just receipts.

Hermes submits a claimed outcome. Your main agent decomposes it into checkable claims and verifies each against the captured transcript.
The hst CLI is one Bun file with zero npm dependencies over the bridge SQLite state. Operators can inspect live tasks, see caller-aware cost snapshots, and export any transcript from the host without adding a service.
Each coding agent plugs into the MCP bridge and routes through the same five operations. Authentication, tool mediation, and auditability sit at the patch bay — not in the client.
