跳到主要内容

withhuman

❖ Communityv1.5.4

Human approval for Hermes tool calls, with human in the loop guardrails. Set rules for what runs automatically, and send the rest to your team to approve or deny from Slack, Discord or your phone. Keep an audit log of every call.

Open in Hermes Desktop
hermes plugins install withhuman

What it adds

Hooks 1

pre_tool_call

README

From the reviewed commit 3fcc9dd ↗; it updates when the author re-pins.

withHuman for Hermes Agent

Human in the loop approvals for Hermes Agent. Let Hermes work on its own, and have a person approve the actions that matter.

withHuman checks every tool call Hermes makes against your approval rules before it runs. Routine calls go straight through. The rest wait for someone on your team to approve or deny them.

You can write rules like:

Let Hermes refund customers up to $20 automatically. Send larger refunds to the support team to approve first.

What you can do

  • Hold the calls that matter. Let reads run, and hold emails, deploys, refunds and record changes for a person.
  • Approve from anywhere. Review requests in the web app, Slack, Discord, email or the withHuman phone app for iPhone or Android.
  • Ask the right people. Send requests to teams, with escalation levels and working hours.
  • Decide routine calls automatically. Approve or deny them with your own webhook, or with an AI judge.
  • Give Hermes your tools. Hermes uses them through the withHuman MCP gateway, without their credentials.
  • Use one set of rules for every agent. The same rules cover Claude Code, Codex, OpenClaw and your own agents.
  • See every tool call. The audit log records every call Hermes makes, and who approved it.

Set up

You need a withHuman organization. Sign up at app.withhuman.ai, or run your own.

hermes plugins install withhuman --enable
hermes withhuman setup

Setup signs you in through your browser and sends a test request. Approve it in withHuman to check everything works.

Then set your approval rules.

Until setup has run, the plugin is off and tool calls run without approval.

Update or remove

hermes plugins update withhuman
hermes plugins remove withhuman

Agent keys expire after 90 days by default. Run hermes withhuman setup again to connect a new one.

Security and footprint

  • Sends: each tool call's name and arguments to your withHuman organization.
  • Stores: the agent key in this profile's secrets (WITHHUMAN_AGENT_TOKEN), and the withHuman address in the plugin's settings.
  • Runs: the withhuman CLI, once for each tool call. It sees only the environment variables it needs, never your provider keys or other plugins' secrets.
  • Downloads: a pinned version of the withHuman CLI from downloads.withhuman.ai, when it is missing or older. The plugin checks it against a checksum it ships with.
  • Changes in Hermes: setup asks to set plugins.hook_callback_timeout to 0, so a call can wait for a person. That setting applies to every plugin's hooks. If you keep the limit, calls that wait for a person longer than it are blocked. On hosted withHuman setup adds the withhuman-gateway MCP server, which sends the agent key to your withHuman address. It offers to install the withhuman and withhuman-pipelines skills from withhuman.ai.
  • Left after removal: the timeout setting, the gateway and the skills. Remove them with hermes config set plugins.hook_callback_timeout 30, hermes mcp remove withhuman-gateway and hermes skills uninstall <name>.
  • Other hooks: Hermes applies argument changes from hooks that run after this one. Keep other pre_tool_call hooks from changing arguments, or an approved call can run with different arguments.

Disclosure

← Back to the catalog · catalog built Nov 3, 2026