Self-hosted · Multi-agent · 24/7

Mission control for a swarm of agents.

Hermes Swarm runs a team of full AI agents on your own hardware — browsing, building, publishing, collaborating peer-to-peer. One real-time console to watch every move and steer the mission.

$ bash <(curl -fsSL https://raw.githubusercontent.com/CyberTron957/hermes-mission-control/main/install.sh)
One line · macOS & Linux · clones, installs, opens the console
or hand it to your AI agent
Paste into Claude Code, Codex, opencode, or Hermes — it installs & configures for you.
[ 01 ]  Deploy

Live in one command.

One line of curl and the swarm is up. It clones, inspects your machine, adopts your provider and keys, and verifies the whole stack before it hands you the dashboard.

  • DETECTSYour machinePython 3.11+, a browser, an existing Hermes config — it picks the right path automatically.
  • ADOPTSYour provider40+ LLM providers via hermes setup. Already configured? It reuses your keys and skips the wizard.
  • VERIFIESThe stackhermes-swarm doctor confirms runtime, model reachability, and Chromium before you start.
bash

[ 02 ]  Anatomy

Every node is a full agent.

Not a prompt in a loop. Each agent has a real terminal, a browser, a filesystem, and the autonomy to use them — then collaborates with the rest of the team to push the mission forward.

Watch them work

Browses the live web

Search and extract from any page. Hits a login, CAPTCHA or 2FA? It hands the browser back to you — embedded right in the dashboard, even on a headless server.

Runs a real terminal

Executes commands, writes and edits files in a shared project directory, and builds its own tools when one is missing.

Collaborates peer-to-peer

Agents message teammates directly on a shared brief. Supervisors review and nudge; a loop detector breaks stalls before they waste a cent.

Publishes to real platforms

Drafts, posts, and ships to live services through the browser — with you approving anything that needs a human sign-off from the inbox.

Tunes itself — with your approval

Agents read their own config and telemetry and propose changes ("raise my iteration limit"). You approve or reject from the inbox.

[ 03 ]  The console

One pane of glass for the whole swarm.

See each agent think, call a tool, and answer — in real time. Read team health at a glance, then intervene with one click: reply to a question, approve a change, or take over a browser.

Hermes·Swarm Live --:--:-- UTC
Network6 agents · 5 links
COORDINATOR RESEARCHER WRITER EDITOR ANALYST QA DESIGNER
Live executionstreaming
09:41COORDassigning weekly digest → 3 agents
09:41RSRCHplanning: gather AI-infra news since Monday…
09:42RSRCHweb_search("ai infrastructure funding")
09:42RSRCH→ 14 sources · extracting top 5
09:43WRITEwrite_file("digest/draft.md")
09:43EDIT
Human input neededWRITER · approve LinkedIn draft before publish
Idle Busy Awaiting human Tool call Blocked
[ 04 ]  Operator in control

Built for unattended runs you can trust.

Daily budget70%
$3.48/ $5.00
content-ops team · resets 00:00 UTC
auto-pauses at the cap — work held, not lost
Schedule3 active
Competitor scan09:00 daily
Metrics pullevery 1h
Monday digestrunning…
cron & one-shot — agents can schedule their own wake-ups
Pending approval1
Researcher proposes
raise max_iterations 90 → 140“hitting the cap on deep-research tasks”
01

Human inbox

Agents ask for a decision, a credential, or a login they can't do alone. It lands in your inbox; you reply and they resume.

02

Daily budget caps

Set a per-team USD ceiling. At the cap, agents pause — work held, not lost — and auto-resume at 00:00 UTC or when you raise it.

03

Embedded browser takeover

Drive an agent's live browser session right inside the dashboard to clear a login or CAPTCHA — then hand it back. Works on a display-less VPS.

04

Scheduled work (cron)

A 9am competitor check, an hourly metrics pull, a Monday digest. Recurring or one-shot — agents can even schedule their own wake-ups.

05

Crash-safe by design

Restart freely. In-flight tasks recover and resume, history persists, and a budget-paused team stays correctly paused.

06

Your keys, your data

Runs entirely on your hardware. Provider keys and credentials stay local on disk with 0600 permissions — never echoed back.

[ 05 ]  How it works

From zero to a working team in three steps.

01

Configure a provider

One hermes setup wizard — pick from 40+ providers, paste a key, choose a model. The swarm reads it directly.

~ 2 MINUTES
02

Build a team

Tell the Architect what you want to accomplish. It proposes the agents, their roles, and who talks to whom — then builds it live on your approval.

GUIDED BY THE ARCHITECT
03

Deploy & steer

Hand the team a mission and watch it run 24/7 from the console. Answer questions, approve changes, adjust budgets — all from one screen.

RUNS UNATTENDED
Architect Building step 02 · team builder
Build a team that ships a weekly AI-infra digest.
Proposed team — 4 agents
Coordinatorplans the brief, assigns work, drives to done
Researcherweb_search + extract across sources
Writerdrafts the digest from findings
Editorfact-checks, tightens, flags claims
[ 06 ]  Self-host

Your hardware. Your rules.

No SaaS, no tenant, no data leaving your box. Run it on a laptop for an afternoon or a VPS for a year — the specifications are the same.

RuntimeDocker or Python 3.11+. The image bundles Python, Hermes, Chromium and the dashboard; agents run contained inside it.
Providers40+ LLM providers via hermes setup — Anthropic, OpenAI, OpenRouter, Groq, DeepSeek, local models — or a custom / OpenAI-compatible endpoint.
Access controlSingle shared API key. Set SWARM_API_KEY and every endpoint and the live WebSocket require it; the dashboard prompts once.
StateEverything on disk under one data dir — team configs, task queues, conversation history, the shared project, and credentials. Rotating backups on every save.
Observability/health + structured logs. Uptime, queue depth, and backend reachability for any uptime monitor; logs to stdout or a rotating file.
LicenseOpen source. Self-hostable, inspectable, yours to fork.

Deploy the swarm tonight.

One clone, one command, one console. Your agents are waiting.

$ bash <(curl -fsSL https://raw.githubusercontent.com/CyberTron957/hermes-mission-control/main/install.sh)