
Hermes Action Bridge
io.github.TheBlueHouse75由 GitHub 账号 TheBlueHouse75 发布v0.6.3更新于 Oct 25, 2026
Delegate real-world actions from Claude Code, Codex, and other agents to Hermes Agent.
概览
本地桥接工具,让编码代理把研究、浏览器流程、定时任务、消息发送等真实操作委派给本机安装的 Hermes Agent。
- 可以做什么
- 把请求委派给 Hermes Agent,可覆盖模式、预设、上下文文件、模型、提供方和超时,也可用快捷工具只请求计划,并通过试运行查看将要发送的确切命令。查看已配置的预设、技能、工具集、运行时与策略,并检查 Hermes 运行时命令。提交有界的异步任务,查询状态、获取结果或取消任务。先准备一个有副作用的操作,再通过一次性短期审批批准或拒绝,批准时客户端会弹出交互式确认。
- 适用场景
- 当在代码仓库中工作的代理需要把编码会话之外的工作——研究、排程、浏览器流程、消息或社交发布——交给已有的 Hermes Agent 设置,而不想重复搭建自动化栈时使用。适合交互式 MCP 客户端;纯自动化环境应改用计划、草稿、试运行或直接命令行。
- 运行要求
- 需要 Node.js 20 或更高版本,并在同一台机器上安装 Hermes Agent 且可通过 hermes 调用(或配置自定义命令路径)。以本地 stdio 进程方式从 npm 安装;可选的 HTTP 服务默认仅监听回环地址,需通过 HERMES_ACTION_HTTP_TOKEN 环境变量提供 bearer 令牌。
安装
在 SourceWeft 中
- 打开 控制台中的 Hermes Action Bridge,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
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.
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 runcommand for one-shot delegation. - Configurable presets for skills, toolsets, provider/model, source, and max turns.
- Deterministic, language-agnostic safety policy: downgrades
executetorequest-approvalby default, with per-preset trust overrides. - Explicit
--yolomode 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:
Installation
Install from npm:
Install the skill and MCP registration for Codex and Claude Code, then verify the complete setup:
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.
From source (for development)
Quick start
Create a config file in your project:
Ask Hermes for a safe plan:
Delegate with a context file:
Run a dry-run to inspect what will be sent to Hermes:
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 usehermes_prepareand thenhermes_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_externalsend_messagesend_emaildeletepaymentgit_pushcredential_change
YOLO mode
YOLO mode is off by default.
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:
Run actions through it — execute stays execute, with no downgrade and no --yolo needed:
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.mdstays 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:
User config path:
Project config path:
Example:
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 crypticE2BIGspawn 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
0600temp file and tells Hermes to read that file (injecting afiletoolset 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 the0600mode 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
Common run options:
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:
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:
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:
npxre-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:
For other JSON MCP clients, or a Claude Code project-scoped configuration,
write or merge .mcp.json directly while preserving other servers:
Or configure it by hand:
Direct delegation
You can always call the bridge directly, without a skill:
MCP tools
Run:
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 effectiveexecuterequest opens an interactive confirmation form first.hermes_plan: shortcut forhermes_runwithmode=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_approveasks 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:
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:
For direct tailnet listening, the bridge accepts only an explicit Tailscale IPv4 address and requires the same bearer-token setup:
The token is never accepted as a CLI value or stored in bridge configuration. Public/wildcard and non-Tailscale addresses are refused.
Development
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
--yoloas 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
来源:README.md,提交 e5a1e6f
工具
0版本历史
1- v0.6.3最新Sep 16, 2026
评分
已满足适用检查项的 50%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分