Install and update
Python 3.10+. Install hermes-gpt from PyPI; verify GitHub and PyPI independently when updating.
Documentation • v0.14.0
Work with evidence. Keep approval in human hands.
Install Hermes GPT, connect a trusted client, and supervise Mission work through declared checks and bounded recovery. These links open the canonical repository guides, which stay the source of truth for current behavior.
Get connected
Private OpenAI access
Use the outbound-only private tunnel path to reach loopback Hermes GPT from supported OpenAI products.
Codex as a client
Configure Codex to use Hermes GPT. Client setup and delegated Codex CLI execution have separate authority gates.
Declare, observe, recover
Autopilot and acceptance
Declare artifact minimum/maximum sizes and optional expected SHA-256 values before dispatch. Requirements survive bounded rework unchanged.
Delivery grace and recovery
Confirmed artifact failures after successful execution get a durable 30-second grace. Recovery proceeds within existing bounds only when every other required check passes.
Flight Deck supervision
Read-only completion checks, fixed failure reasons, delivery grace, and pending validation. Changing or unreadable evidence remains unverified.
Understand the boundaries
Operator and Owner policy
Default read-only behavior, dry-run-first mutations, explicit authority, and secret-path protections. Autopilot remains default off.
Missions and approval
Autopilot stops at approval frontiers. Final Mission approval remains Owner-only; size and integrity checks do not judge semantic quality.
Authentication and compatibility
Configure bearer or OAuth authentication for trusted clients. Public unauthenticated Operator hosting is unsupported.