Hermes Action Bridge is a configurable bridging tool used to let external agents delegate real-world operations to the Hermes Agent without needing to reimplement Hermes's skills, tools, platform integrations, browser automation, scheduled tasks, or messaging flows. It can be used from Claude Code, Codex, Cursor, CI jobs, shell scripts, or other MCP-capable clients. MCP execution requires interactive form confirmation and fails closed in non-interactive clients; planning, drafting, dry runs, and direct CLI workflows remain available for automation.
Main capabilities: provides a generic hermes-action run command for one-off delegation; supports configurable presets for skills, tool sets, providers/models, sources, and maximum turns; has a deterministic, language-agnostic safety policy that by default downgrades execute to request-approval and supports per-preset trust overrides; provides an explicit --yolo mode to bypass bridge-level policy; supports context file injection with per-file limits and a configurable total budget; automatically passes large contexts, with oversized prompts handed to Hermes via secure temporary files; has a timeout limit per run; dry-run mode for debugging the exact Hermes command, prompt, and computed prompt size; provides complete Codex and Claude Code installers; supports MCP capability discovery, cancellable asynchronous jobs, and one-time two-phase approval; provides local stdio transport as well as an optional, authenticated Streamable HTTP transport for Tailscale-restricted hosts.
Usage: requires Node.js 20 or higher, and Hermes Agent installed and available. It can be installed via npm, use the installer to install the skill and register MCP for Codex and Claude Code, then verify the setup. CLI commands can also be used directly to create configuration files, request plans, delegate context files, run dry runs, etc. Execution modes include plan (returns only a plan), draft (generates artifacts only, with no external side effects), execute (may perform permitted operations, still following Hermes's own safety rules), and request-approval (prepares operations but does not execute them; MCP callers should use hermes_prepare and hermes_approve, which opens an interactive client form showing the exact operation and executes only after human confirmation). The policy guard is deterministic and language-agnostic: in execute mode, the bridge switches to request-approval by default unless the preset is explicitly trusted or --yolo is set. Every valid MCP execute request requires interactive client confirmation, including trusted presets and YOLO. The bridge does not attempt to infer risk from keywords.
Use cases: suitable for coding agents or other automation flows that need to delegate operations such as research, scheduling jobs, opening browser workflows, sending messages, preparing social posts, using Hermes skills, or coordinating with messaging gateways to Hermes. Supports invocation from Claude Code, Codex, Cursor, CI jobs, shell scripts, etc.
Limitations: MCP execution requires interactive form confirmation and fails closed in non-interactive clients. YOLO mode is off by default and does not bypass the MCP interactive confirmation barrier, Hermes Agent's own safety rules, provider/tool approval prompts, or platform constraints. Direct CLI execution assumes use by a trusted, supervised operator. The MCP specification requires clients to provide UI, identify the requesting server, and provide deny/cancel options; malicious clients or hooks configured to auto-answer can forge confirmations. Do not auto-answer MCP confirmations, do not add hermes-action run (execute/yolo) to an allowlist, and do not put secrets in .hermes-action.yaml.
