Docs / Hermes Agent

Hermes Agent

Hermes is the persistent background worker tier — an always-on, self-improving MCP bridge that runs 24/7 on the AI OS platform. While Claude is the "brain" (planning and decisions), Hermes is the "hands" (interacting with the real world around the clock).

Architecture

Hermes connects to AI OS as a Model Context Protocol (MCP) server. The Orchestrator delegates tasks to Hermes for persistent, background execution:

  • Claude (AI OS) — Planning, decision-making, orchestration, complex reasoning
  • Hermes — 24/7 execution, mobile bridge, cron jobs, approval gating, long-running tasks

The MCP connection is managed by the platform and configured automatically for each tenant.

Remote Approval Gate

When Hermes encounters a destructive or risky operation, it pauses execution and pushes an approval prompt to your phone:

  1. A task requests a risky action (file deletion, force-push, database modification)
  2. Hermes suspends the action and creates an approval request
  3. A notification is sent via Telegram, Slack, or Discord
  4. You approve or reject from your phone
  5. Hermes resumes or aborts based on your decision

Approval requests are tagged with risk levels:

Risk LevelExamplesDefault Action
HighForce-push, database drop, production deploy, key rotationBlock until approved
MediumFile deletion, config changes, branch deletionBlock until approved
LowLog cleanup, cache purge, temp file removalAuto-approve after 5 min

Walkaway Mode

Initiate a long-running task from the dashboard, then close your laptop. Hermes acts as the bridge:

  • Progress pings — Sends status updates to your phone at each milestone
  • Mobile replies — Reply from your phone with instructions: "keep going," "make it simpler," "push to GitHub"
  • Completion notification — Alerts you when the task finishes with a summary
  • Approval integration — If the task hits a risky action, you get an approval prompt on mobile
Example: Start a massive refactoring task, walk away, and Hermes will ping you on Telegram with progress updates. Reply "push to staging" from your phone when it's done.

Remote Bug Diagnosis

When a bug report arrives in Discord or Slack, use Hermes from your phone to trigger analysis:

  1. Send a message to Hermes via Telegram: "Analyze the auth bug reported in #bugs"
  2. Hermes reads the bug report, instructs Claude to analyze the relevant files
  3. Claude's diagnosis is sent back to your phone
  4. Optionally, approve a fix from mobile

Scheduled Skills

Hermes has six named skills. Three run for real today — dispatched to a real agent on a schedule, or on demand from the dashboard or via /api/hermes/delegate. The other three are still simulated pending their own real integrations.

Real skills

JobScheduleDescription
Dev ProjectOn demandGrok Build plans a platform upgrade — see Self-Improve
News Brief0 6 * * *Live intelligence sweep via the scout agent
Uptime Check*/30 * * * *Health-snapshot review via the sysadmin agent (real disk, memory, and load metrics)

Simulated skills

The schedule and dashboard UI exist for these, but delegating them does not yet call a real agent or touch a real external service — treat their output as a placeholder, not a real result:

JobScheduleWhy it's not real yet
Inbox Summary0 8 * * *No email/OAuth integration yet
GitHub Backup0 2 * * *Scope not yet decided (state snapshot vs. repo mirror)
Comment Monitor0 */4 * * *No YouTube/social polling integration yet

Platform Self-Improvement

Hermes's Dev Project skill hands off to Self-Improve, which reads AI OS's own code and proposes upgrade plans using xAI's Grok Build — either to apply to your instance or as a draft pull request against the public distribution repo. Every apply and every PR always requires your explicit approval; see the Self-Improve docs for the full safety architecture.

Frequently asked questions

Is Hermes a separate server I need to run?

No. Hermes's real skills run as an in-process dispatcher inside your AI OS instance — there is no standalone Hermes MCP server to install or manage.

How do I know if a Hermes skill is real or simulated?

Check the tables above, or call GET /api/hermes/status — each entry in the returned skills array is tagged real: true when it dispatches to an actual agent.

How is Hermes's Dev Project skill different from Self-Improve?

They're the same underlying feature reached two ways — Dev Project is Self-Improve's planning flow triggered through Hermes's delegate API instead of the dedicated Self-Improve dashboard view. See Self-Improve for the full picture.

API Endpoints

GET /api/hermes/status

Check MCP connection status, uptime, stats, and loaded skills.

POST /api/hermes/delegate

Delegate a task to Hermes. Modes: background, walkaway, cron (simulated), or the real dev-project, news-brief, uptime-check.

GET /api/hermes/tasks

List all active delegated tasks with progress and logs.

GET /api/hermes/approvals

Get pending approval requests with risk levels and context.

POST /api/hermes/approvals/:id

Approve or reject a pending action. Body: { "decision": "approve" }

GET /api/hermes/cron

List all scheduled Hermes cron jobs.

POST /api/hermes/cron

Create a new Hermes cron job.

Configuration

VariableDefaultDescription
HERMES_MCP_URLhttp://127.0.0.1:8420MCP server endpoint for Hermes
TELEGRAM_BOT_TOKENRequired for mobile approval prompts via Telegram
SLACK_WEBHOOK_URLRequired for Slack notifications from Hermes

Key Benefits

  • Tenant isolation — Each tenant's Hermes tasks are fully isolated from other tenants.
  • Cost efficiency — Background and cron tasks use your own API keys at standard rates.
  • Always available — Unlike interactive sessions, Hermes works 24/7 without requiring your browser.