Hermes LoopHermes Loop
System

How Hermes Loop works

Nine surfaces, one flow: messy human request → governed agent run → provable receipt. Every node links to its live page; counts reflect the current workspace.

Guided demo →Dashboard →Hermes Agent parity →Hermes · onlineHermes memory · online
Plain English

The control room for the Hermes Agent.

Hermes Agent is Nous Research's open-source autonomous agent — it lives on your server, remembers what it learns, and gets more capable the longer it runs. Hermes Loop is the operator surface: you launch missions, run named crews of subagents, gate risky outputs with approvals, and produce a hashed receipt that proves what happened.

What Hermes Agent gives you

The autonomous engine: persistent memory, subagents with their own conversations and terminals, natural-language cron, and native tools (web search, browser automation, vision, multi-model reasoning).

What Hermes Loop adds on top

The governance layer: crews, a job queue, approvals, hashed receipts, a trust ledger, evals, schema self-correction, real-cost accounting, and a full audit trail.

Why it is different

A chatbot gives an answer. The Hermes Agent does the work autonomously; Hermes Loop proves it: who ran, which tools were called, which memory was used, what was approved, and what hash signed off the run.

What to try first

Run Bug Hunter on /demo-target, approve the QA report, generate the receipt, then open Trust to see the run roll up into risk and proof.

How this app relates to Hermes Agent

Hermes Agent is the engine. Hermes Loop is the control room.

Every Hermes Agent capability has a corresponding surface in this app that governs, audits, and proves it. Hermes Agent gives you autonomy; Hermes Loop gives you governance and proof.

Hermes Agent (the engine)
Hermes Loop (this app)
Isolated subagents with their own conversations + terminals
Crews — named, ordered, Zod-validated subagent sequences.
Natural-language cron scheduling
/schedules + the worker process running RUN_MISSION jobs.
Persistent learning memory
/memory — operator-approved, audited, diffable, hygiene-checked.
Native tools (web search, browser automation, vision, TTS, multi-model reasoning)
Sandboxed tool layer with approval gates and call logs.
Multi-platform interface (Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI)
A single operator-grade web surface that proves what happened.
Execution backends (local, Docker, SSH, Singularity, Modal)
Backend-agnostic — the worker process just runs missions; deploy where you like.
(autonomy)
Approvals, hashed receipts, trust ledger, evals harness — governance + proof.
Mission walkthrough

What happens when you run a mission

1. Request lands
You type an objective (or drop one in the inbox; triage proposes a crew + objective).
2. Crew is selected
A built-in or custom crew runs, agent by agent, with Zod-validated outputs at each step.
3. Job is queued
The worker picks up the RUN_MISSION job, with retries and exponential backoff on transient failures.
4. Subagents run with tools + memory
Each subagent calls Hermes, can invoke sandboxed tools (browser_qa_audit, web_snapshot, document_extract…), and is fed approved memory from the vault.
5. Risky outputs wait for approval
Drafts, trade tickets, exports, gated tools — nothing leaves the desk without your decision.
6. Receipt + trust
Every prompt, response, tool call, and approval persists. A WorkflowReceipt with an integrity hash rolls up into the trust ledger.
Comparison

Normal chatbot vs Hermes Loop

Hermes (or any LLM) gives you reasoning. Hermes Loop wraps that reasoning with the controls a real workflow needs.

Axis
Normal chatbot
Hermes Loop
Output
One answer.
Multi-step mission with every step saved.
Tools
Usually hidden or ad hoc.
Tool calls logged, schema-validated, sandboxed.
Memory
Implicit context.
Operator-approved memory with usage records and diffs.
Safety
Model decides.
Human approval gates risky outputs.
Proof
Transcript.
Receipt with integrity hash, timeline, risk score, trust ledger.
The nine surfaces
011 new · 3 triaged
Inbox
Triage layer. Raw requests from desktop or mobile become structured missions.
02deterministic + Hermes
Triage Agent
Classifies the request, proposes crew + objective + cadence, surfaces missing info.
030 queued · 0 running · 0 failed
Jobs
Persistent queue. Triage, mission runs, schedule fires, resume-after-approval. Retryable + observable.
044 active · 10 total
Missions
A mission is one unit of work. A crew of specialist agents executes it step by step with Zod-validated outputs.
0514 tool calls · 9 memories
Tools + Memory
Sandboxed tools called inline. Approved memories injected as context. Both fully logged.
065 pending
Approvals
User decision gate. Drafts, trade tickets, exports, follow-ups, gated tools — nothing leaves the desk on its own.
071 active
Schedules
Cron-like cadences materialize fresh missions and dispatch them through the same orchestrator.
084 total · 2 risky
Receipts
Black-box recorder per mission. sha256 integrity hash. Markdown + JSON export. Finalize to lock.
09live signals
Trust ledger
Aggregate reliability across the whole desk: approval rate, tool failures, hygiene flags, Hermes sync.
Common workflows

What people actually do here

Audit a website
  1. 1Open Bug Hunter
  2. 2Launch with the target URL
  3. 3Wait for the run
  4. 4Approve the QA report
  5. 5Open the receipt
Drop a messy request
  1. 1Paste any text in the inbox
  2. 2Wait for triage to finish
  3. 3Review the proposed crew + objective
  4. 4Approve & convert into a mission
Schedule recurring work
  1. 1Open Schedules
  2. 2Pick a crew + cadence (daily/weekly/once)
  3. 3Worker fires it on time
  4. 4Each fire becomes a mission with a receipt
Create a custom crew
  1. 1Open the crew builder
  2. 2Add 2–6 agents with roles + JSON schemas
  3. 3Save the crew template
  4. 4Launch a mission with it
Prove a run with a receipt
  1. 1Find the mission
  2. 2Generate or open its receipt
  3. 3Inspect the timeline + integrity hash
  4. 4Finalize to lock; export markdown/JSON
  5. 5Risk rolls up to the trust ledger
Vocabulary (use these words)
MissionOne run of a crew end-to-end.
CrewAn ordered sequence of specialist agents.
JobA background work item in the worker queue.
ApprovalA user decision gate on a risky output (draft, trade, export, gated tool).
MemoryOperator-approved context injected into agent prompts.
ReceiptHashed black-box recording of one mission run.
Trust ledgerAggregated reliability roll-up read from receipts.
ToolA sandboxed action agents can call (web_snapshot, browser_qa_audit, etc).
Operator surfaces
Hermes control room →Skill marketplace →Tool sandbox →Memory inspector →Crew library →Memory diff →Memory hygiene →Settings + demo data →