Hermes Agent features,
with proof and control.
Run Hermes-powered crews that browse, search, inspect a codebase, call Python, and analyze media. Hermes Loop wraps every run in a triage inbox, a persistent job queue, approval gates, and a hashed receipt that proves what happened.
- Learning loop — every settled mission distils into reusable Skill rows on the same crew.
- Cross-session memory recall — operator-approved context queried across all prior runs, injected into the next one.
- MCP integration — native Model Context Protocol client; remote tools surface alongside built-ins.
Not just a dashboard. Real agent capabilities with receipts.
Hermes does the reasoning. Hermes Loop wires the runtime around it: browser, terminal, Python, search/media providers, schedules, webhooks, approvals, memory, evals, and proof.
Playwright QA audits URLs with screenshots, console capture, accessibility and layout checks.
terminal_exec runs safe repo diagnostics and approval-gates risky commands.
python_rpc runs short scripts in an ephemeral workspace with policy checks.
web_search live via Tavily, with Brave + SerpAPI as fallback providers.
vision_analyze on Gemini 2.5 Flash via Hermes — image URL or base64 → structured findings.
image_generate uses Hermes image-capable models (Gemini, Flux). text_to_speech uses ElevenLabs eleven_multilingual_v2.
Hermes parses plain-English schedules into the existing cadence engine.
Signed inbound webhook turns external messages into InboxItems and triage jobs.
Per-role routing wired: HERMES_MODEL_FAST / STRONG / JUDGE / VISION. Unset roles fall back to default; receipt records fallbackUsed=true.
After every mission settles, Hermes Loop distils up to 3 reusable lessons into Skill rows. Future runs of the same crew inject them automatically.
Operator-approved memory queried across all prior sessions, top matches injected into the Triage Agent + first-step prompts.
Native MCP (Model Context Protocol) client. Configure MCP_SERVERS and remote tools surface alongside built-ins.
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.
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).
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.
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.
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.
What people can actually do with it
Bug Hunter opens the page in Playwright, captures evidence, finds accessibility/conversion issues, and stages a QA report for approval.
Codebase Debugger lets Hermes use governed terminal commands, then writes a diagnosis without editing files.
Provider-backed web search, vision, image generation, and TTS are testable from the Media section before agents use them.
Use natural-language schedules or a signed webhook to turn external requests into queued agent missions.
Define agents, roles, instructions, tools, and execution order, then run that crew like a built-in workflow.
Receipts and Trust show the timeline, tools, approvals, memory, risk, model cost source, and integrity hash.
Crews that
actually work.
Three pre-built crews ship in the box. Each is a sequence of specialists with strict output schemas and a sample deliverable you can preview right now.
Paper Trading Desk
Research, thesis, and a paper-trade ticket — fully simulated.
Bug Hunter Crew
Crawl, test flows, audit a11y, and ship a client-ready report.
Life Admin Crew
Evidence, policy angle, draft, and a follow-up plan.
Codebase Debugger Crew
Repo scout · build runner · error analyst · fix planner · report.
From prompt to
completed mission.
Six explicit stages. Each one writes to the audit trail. Nothing jumps a stage; nothing leaves the desk on its own.
Tools agents can
actually call.
Agents request tools by name. The desk validates input, runs the tool, persists the call, and feeds the result into the next prompt. Every tool is sandboxed — no shell, no live trading, no auto-send.
Fetches a public URL — title, description, headings, sample links, text sample. Blocks localhost, private IPs, file URLs, and metadata endpoints. 8s timeout, 1 MB body cap.
Returns deterministic synthetic OHLCV — never real broker data. Used by the Backtest Agent to compute win rate and drawdown.
Pulls dates, amounts, names, companies, order numbers, and key claims from mission-owned documents only.
Creates an approvable follow-up reminder. Never sends an email or schedules an external action.
Renders a deliverable to MARKDOWN or JSON. Approval-gated by default — the export is staged, not shared.
- · No shell execution
- · No browser form submission
- · No live trading or brokerage APIs
- · No outbound emails or messages
- · web_snapshot blocks localhost + private IPs
- · Tool I/O is persisted for audit
Persistent by design.
A chat window forgets. A desk remembers. Schedules, replays, run history, and the approval inbox give you continuity instead of a transcript.
Daily, weekdays, weekly, or monthly cadences. Each fire materializes a fresh mission and runs the orchestrator.
Animate the agent + tool timeline of any completed mission. Useful for review, demos, and onboarding.
Every step is keyed to its mission. Tokens, latency, prompts, raw responses, parsed outputs — all retained.
Drafts, trade tickets, exports, reminders, and tool gates land here. Approve, reject, or request changes.
Built for control.
Auditability isn't a setting — it's the product. Every prompt is on disk. Every tool call is on disk. Every risky action waits for a human signature. Every deliverable traces back to the agent step that made it.
- Raw prompt + raw response stored on every step
- Zod schema validation on every parsed output
- Tool input + output persisted with timing + status
- Approval queue for trades, drafts, exports, follow-ups, tool gates
- Per-mission audit log of agent + tool + system events
- Demo mode runs the full pipeline with no external calls
Your operations terminal,
running locally.
Spin up the dashboard, pick a crew, type an objective, and watch six agents work — calling tools, queuing approvals, writing audit events, shipping a deliverable. Plug in Hermes for live runs or stay in demo mode — the interface never changes.
- 1Open /dashboard
- 2Pick a crew (or seed the demo workspace)
- 3Type an objective
- 4Click Run mission
- 5Watch agents + tools execute
- 6Approve the queued items