Hermes Action Bridge

Hermes Action Bridge

A configurable bridge that lets external agents delegate real-world actions to Hermes Agent without reimplementing Hermes skills, tools, platform integrations, browser automation, cron jobs, or messaging flows.

Use it from Claude Code, Codex, Cursor, CI jobs, shell scripts, or another MCP-capable client. MCP execution requires interactive form elicitation and fails closed in non-interactive clients; plan, draft, dry-run, and direct CLI workflows remain available to automation.

external agent -> hermes-action -> Hermes Agent -> skills/tools/integrations

Why this exists

Coding agents are good at understanding a repository. They should not duplicate your automation stack.

If an agent needs to do something outside its local coding session — research, schedule a job, open a browser workflow, send a message, prepare a social post, use a Hermes skill, or coordinate with a messaging gateway — it can delegate that request to Hermes through this bridge.

Features

• Generic hermes-action run command for one-shot delegation.
• Configurable presets for skills, toolsets, provider/model, source, and max turns.
• Deterministic, language-agnostic safety policy: downgrades execute to request-approval by default, with per-preset trust overrides.
• Explicit --yolo mode for users who intentionally want to bypass bridge-level policy.
• Context file injection with a per-file cap and a configurable aggregate budget (clear error instead of a cryptic E2BIG).
• Automatic large-context delivery: an oversized prompt is handed to Hermes through a secure temp file instead of overflowing the command line.
• Per-run timeouts (configurable, with per-mode defaults) that always reap the Hermes child process.
• Dry-run mode for debugging the exact Hermes command, prompt, and computed prompt size.
• Complete Codex and Claude Code installer: managed skill plus verified global MCP registration.
• MCP capability discovery, cancellable asynchronous jobs, and one-shot two-phase approvals gated on interactive form elicitation, with a privacy-preserving local audit log.
• Local stdio transport plus an opt-in, authenticated Streamable HTTP transport for Tailscale-restricted hosts.
• No project-specific assumptions. All behavior is configured through YAML and CLI flags.

Requirements

• Node.js 20 or newer.
• Hermes Agent installed and available as hermes, or configured with a custom command path.

Check Hermes:

hermes --version

Installation

Install from npm:

npm install -g hermes-action-bridge

Install the skill and MCP registration for Codex and Claude Code, then verify the complete setup:

hermes-action --version
hermes-action install all
hermes-action doctor

Global installation is all-or-nothing for the requested agents. If Codex or Claude Code is unavailable, conflicting, or cannot verify its user-scoped stdio MCP entry, the command exits non-zero without installing the requested skills. Use --project for an explicit skill-only setup.

Other distribution channels

Choose the channel that matches the agent surface. Hermes Agent must always be installed and configured on the same computer as the local bridge.

The standalone Agent Skill only teaches an agent when and how to delegate; it does not register an MCP server. Use the npm installer, Claude marketplace, or MCPB channel when MCP tools are required.

Maintainers releasing a new version should follow the distribution runbook, which covers the automated release workflow and the manual directory publication checks.

Codex CLI and the Codex surface in the ChatGPT desktop app share the local Codex MCP configuration on the same host. A general ChatGPT connector is a different distribution target: it requires a remotely reachable MCP server and is not installed by this local stdio bundle.

git clone https://github.com/TheBlueHouse75/hermes-action-bridge.git
cd hermes-action-bridge
npm install
npm run build
npm link

Quick start

Create a config file in your project:

hermes-action init

Ask Hermes for a safe plan:

hermes-action run --mode plan "Find the best next action from this repository context."

Delegate with a context file:

hermes-action run \
  --preset research \
  --context ./notes.md \
  "Analyze this and return the next concrete action."

Run a dry-run to inspect what will be sent to Hermes:

hermes-action run --dry-run --json "Summarize this project."

Execution modes

• plan: Hermes is instructed to return a plan only. Treat it as open-world because configured tools remain available.
• draft: Hermes produces an artifact only. No external side effects.
• execute: Hermes may execute allowed actions, while still following Hermes' own safety rules.
• request-approval: Hermes prepares an action without executing it. MCP callers should use hermes_prepare and then hermes_approve; the latter opens an interactive client form showing the exact action and executes only after the human confirms it.

The policy guard is deterministic and language-agnostic: in execute mode the bridge switches to request-approval by default — regardless of the prompt's wording or language — unless the preset is explicitly trusted (empty require_approval_for) or --yolo is set. Separately, every effective MCP execute request requires interactive client confirmation, including trusted presets and YOLO. The bridge does not try to infer risk from keywords.

Risk categories below are informational only: the bridge surfaces the ones it recognizes in the prompt envelope to help you and Hermes decide, but they never drive the mode. (Keyword matching is English-biased and trivially evaded, so it must not be a security control.)

• publish_external
• send_message
• send_email
• delete
• payment
• git_push
• credential_change

YOLO mode

YOLO mode is off by default.

hermes-action run --yolo --mode execute "Do the task now."

YOLO bypasses the mode-downgrade policy. It does not bypass the MCP interactive confirmation barrier, Hermes Agent's own safety rules, provider/tool approval prompts, or platform constraints. Direct CLI use remains an explicit trusted-local escape hatch.

Use it only when the caller and environment are trusted.

Trusted presets (direct execution)

By default every execute is downgraded to request-approval — the right default when an agent might act unattended. When you are the one asking for an action and want it to run without that extra gate, define a trusted preset: one whose require_approval_for is an empty list. Put it in your user config so it is yours alone and available in every project:

# ~/.config/hermes-action/config.yaml   (user scope, merged before any project config)
presets:
  act:
    description: Direct execution of an action I explicitly asked for (trusted).
    require_approval_for: []

Run actions through it — execute stays execute, with no downgrade and no --yolo needed:

hermes-action run --preset act --mode execute "post the release note to #general"

Keep it safe:

• A trusted preset relaxes the mode-downgrade policy. Effective MCP execution still requires interactive confirmation; direct CLI execution relies on the operator and environment being trusted.
• Keep it in your user config; it never ships in the public package, and the distributed SKILL.md stays cautious for everyone else.
• Reserve trusted presets for actions a human explicitly asks for. Leave the conservative default for anything an agent could trigger on its own.

Configuration

hermes-action loads config in this order:

CLI flags > project .hermes-action.yaml > user config > built-in defaults

User config path:

~/.config/hermes-action/config.yaml

Project config path:

.hermes-action.yaml

Example:

runtime:
  adapter: hermes-cli
  command: hermes
  # Aggregate budget across all --context files, in bytes (default ~768 KiB).
  # Raising this above ~896 KiB switches large prompts to temp-file delivery.
  max_context_bytes: 786432
  # Optional overall per-run timeout in seconds. When unset, per-mode defaults
  # apply: 180s for plan/draft, 600s for execute/request-approval.
  # timeout_seconds: 600

defaults:
  mode: plan
  source: external-agent
  max_turns: 30
  preset: default

presets:
  default:
    description: No extra skills or toolsets. Uses the active Hermes profile.
    skills: []
    toolsets: []

  research:
    description: General research and synthesis.
    skills: []
    toolsets: [web, terminal, file]
    # Per-preset override: relax approvals for a trusted preset.
    # An empty list never downgrades execute to request-approval.
    # require_approval_for: []

  coding:
    description: Repository inspection and runtime validation.
    skills: [developer-assurance-and-validation, runtime-debugging]
    toolsets: [terminal, file]

policy:
  yolo: false
  # Global default; a preset's own require_approval_for takes precedence when set.
  require_approval_for:
    - publish_external
    - send_message
    - send_email
    - delete
    - payment
    - git_push
    - credential_change

Limits, timeouts, and large context

The bridge builds one prompt envelope (policy header + your request + every <context> block) and hands it to Hermes. Two guardrails keep that reliable:

• Aggregate context budget (runtime.max_context_bytes, default ~768 KiB, plus a fixed 250 KiB per-file cap). Exceeding it fails fast with an actionable message — Context total N bytes exceeds the limit ... Split your handoff or raise runtime.max_context_bytes. — instead of a cryptic E2BIG spawn crash. The default leaves headroom under the operating system's argument-size limit.
• Large-context delivery. If you raise the budget and the envelope grows past ~896 KiB, the adapter writes it to a 0600 temp file and tells Hermes to read that file (injecting a file toolset for the call). This sidesteps the OS argument limit. Two caveats: it adds one tool-calling turn, and the real ceiling becomes the model's context window — a multi-megabyte file may only be partially read. The temp file is removed after the run; only an uncatchable kill of the bridge process itself could leave it behind (mitigated by the 0600 mode and the OS temp reaper).

Every run is bounded by a timeout (default 180s for plan/draft, 600s for execute/request-approval). Override per run with --timeout <seconds> or globally with runtime.timeout_seconds. On expiry the Hermes child is sent SIGTERM, then SIGKILL after a short grace, and the result is marked as timed out.

hermes-action doctor prints the effective limits, and hermes-action run --dry-run --json reports the computed prompt size (promptBytes, promptChars) and whether delivery would use argv or a temp-file.

CLI reference

hermes-action init [--file .hermes-action.yaml] [--force]
hermes-action run [options] "request"
hermes-action presets [--json]
hermes-action status [--json]
hermes-action mcp
hermes-action serve [--listen 127.0.0.1] [--port 8765]
hermes-action install <claude-code|codex|all|mcp> [options]
hermes-action uninstall <claude-code|codex|all|mcp> [options]
hermes-action doctor [--json] [--probe]

Common run options:

--mode <plan|draft|execute|request-approval>
--preset <name>
--context <path...>
--config <path>
--provider <name>
--model <name>
--max-turns <number>
--timeout <seconds>
--source <name>
--yolo
--dry-run
--json

Native agent skills

Instead of pasting instructions and MCP snippets by hand, run the complete installer. It installs the open-standard skill and registers the local stdio MCP server through each agent's own CLI:

hermes-action install all            # skills + global MCP for both agents
hermes-action doctor                 # verifies skills, registrations, and MCP handshake
hermes-action install claude-code --print     # preview, write nothing
hermes-action install codex --project-hint     # also add a marker block to AGENTS.md
hermes-action uninstall all

Install behavior is safe by default: it uses a stable absolute bridge path, does not depend on an interactive PATH, never modifies CLAUDE.md / AGENTS.md unless you pass --project-hint, and refuses to overwrite either a foreign skill or a customized MCP entry. Re-running it is idempotent. Use --project for a project-local skill only and --dry-run to preview operations.

The generated skill is shown in examples/claude-code/SKILL.md. Project-hint usage is documented in examples/claude-code/CLAUDE.md and examples/codex/AGENTS.md.

Codex plugin (one-line install)

Codex users can install the delegation skill and the MCP server together as a plugin:

codex plugin marketplace add TheBlueHouse75/hermes-action-bridge
codex plugin add hermes-action@hermes-action-bridge

This registers the hermes-action MCP server and the Hermes delegation skill in Codex. Hermes Agent must be installed locally — the plugin runs the version-pinned npm package with npx.

Faster MCP startup: npx re-resolves the package on every server start. The native installer registers the absolute global launcher instead.

MCP configuration

Register the global MCP server for Codex and Claude Code without reinstalling their skills:

hermes-action install mcp

For other JSON MCP clients, or a Claude Code project-scoped configuration, write or merge .mcp.json directly while preserving other servers:

hermes-action install mcp --write

Or configure it by hand:

{
  "mcpServers": {
    "hermes-action": {
      "command": "hermes-action",
      "args": ["mcp"]
    }
  }
}

Direct delegation

You can always call the bridge directly, without a skill:

hermes-action run --mode plan "Ask Hermes what should happen next."

hermes-action run \
  --preset coding \
  --context ./codex-notes.md \
  "Use Hermes to validate this plan and identify missing runtime checks."

MCP tools

Run:

hermes-action mcp

Exposed tools:

• hermes_run: delegate a request to Hermes. Optional overrides: mode, preset, contextFiles, yolo, dryRun, and — to trade cost for speed on simple tasks — model, provider, maxTurns, timeoutSeconds. Any effective execute request opens an interactive confirmation form first.
• hermes_plan: shortcut for hermes_run with mode=plan.
• hermes_capabilities: report configured presets, skills, toolsets, runtime, and policy without inventing live integrations.
• hermes_presets: list configured presets.
• hermes_status: check the configured Hermes runtime command.
• hermes_submit, hermes_job_status, hermes_result, hermes_cancel: manage bounded, process-local asynchronous jobs.
• hermes_prepare, hermes_approve, hermes_reject: create and consume a short-lived one-shot approval. hermes_approve asks the MCP client for interactive confirmation before execution.

The MCP surface delegates to Hermes instead of mirroring every Hermes tool. Jobs and approvals are intentionally process-local: restarting the MCP server clears them. Audit events are written as owner-only JSONL metadata under $XDG_STATE_HOME/hermes-action/audit.jsonl (or ~/.local/state/hermes-action/audit.jsonl) without prompt, context, token, or full result content.

Remote MCP over Tailscale

The HTTP transport is opt-in and loopback-only by default:

hermes-action serve --listen 127.0.0.1 --port 8765

Prefer placing Tailscale Serve in front of that loopback endpoint. Forwarded *.ts.net requests still require a bearer token, so start the loopback server with --token-env:

export HERMES_ACTION_HTTP_TOKEN="<at-least-32-random-bytes>"
hermes-action serve \
  --listen 127.0.0.1 \
  --port 8765 \
  --token-env HERMES_ACTION_HTTP_TOKEN

For direct tailnet listening, the bridge accepts only an explicit Tailscale IPv4 address and requires the same bearer-token setup:

hermes-action serve \
  --listen 100.64.0.10 \
  --port 8765 \
  --allow-tailnet \
  --token-env HERMES_ACTION_HTTP_TOKEN

The token is never accepted as a CLI value or stored in bridge configuration. Public/wildcard and non-Tailscale addresses are refused.

Development

npm install
npm run build
npm run test
npm run check

Functional tests use a fake Hermes binary to verify command construction, prompt wrapping, risk policy, and YOLO behavior without spending LLM credits.

Documentation

• Architecture: module map, config precedence, modes, policy, and MCP design.
• Functional testing: automated tests and live smoke-test procedures.
• Contributing: local setup, development rules, and release checklist.
• Changelog: release history.
• Security policy: how to report a vulnerability.

Security model

Direct CLI execution assumes a trusted, supervised operator. MCP execution has an additional protocol-enforced gate: every effective execute request asks the client for MCP form elicitation, shows the exact action and metadata, and runs only when the response is both accept and confirm=true.

For side effects, prefer the testable two-phase flow: hermes_prepare returns a local no-tool summary and opaque approval ID, and hermes_approve opens the interactive confirmation before consuming the unchanged request exactly once. Unsupported clients, CI jobs, shell scripts, cancellations, timeouts, and declines fail closed without execution. Use plan, draft, dry-run, or the direct CLI when an interactive MCP client is unavailable; do not describe the direct CLI as providing a distinct human-approval barrier.

This gate assumes a conforming, trusted MCP client. The MCP specification requires clients that advertise elicitation to provide UI, identify the requesting server, offer decline/cancel, and let the user review the response. A malicious client, or an elicitation hook configured to auto-answer, can forge accept and confirm=true; do not expose the bridge to untrusted MCP clients. The bridge does not claim a human boundary in that deployment model.

The bridge's own guard is a secondary net: execute is downgraded to request-approval by default unless a preset is trusted (empty require_approval_for) or --yolo is set. The bridge never injects --yolo or a bypass flag on its own.

• Do not auto-answer MCP elicitation or install an elicitation hook that accepts execution forms. Doing so deliberately removes the human boundary.
• Do not allowlist hermes-action run (execute/yolo). A command-pattern allowlist cannot distinguish a safe plan from an execution with side effects.
• Do not put secrets in .hermes-action.yaml. Keep provider credentials in Hermes Agent, your OS keychain, or the platform's secure store.
• Treat --yolo as a trusted-local escape hatch, not a default.

Also by the author

Built by Cyril Guilleminot, who also makes offline-first voice tools:

• Weesper Neon Flow — speak text into any app at 3× typing speed, fully offline (macOS & Windows).
• Weesper Transcribe — offline transcription on the macOS App Store.

License

MIT