AGENT MACHINES
stateful microVMper-account fleetPersistent Machinesfor your Agent
One stateful microVM per account. Boot in 30 seconds, sleep on idle, wake on the first prompt. memory . cron . sessions . MCP-native.
State on disk, not in RAM
Chats, files, memory, sessions, crons -- all in /home/machine. Survives every sleep.
Per-account fleet
One Clerk sign-in. Same machine on every device. Keys + agent choice in private metadata.
Bring any agent + tool
Hermes or OpenClaw. 95 skills, 23 built-ins, 17 services. Dedalus live; Sandbox + Fly shaped.
ACTIVITY -- 6 MONTHS
116 active dayshover or click a cell
services touched . 20
click to filter
MACHINE -- STACK
vCPU
1
second-billed
memory
2 GiB
storage
10 GiB
persists across sleeps
boot
<30s
cold . <5s warm
skills
95
bundled + wiki
fleet
per-account
Clerk-tied
CAPABILITIES
Machine first. Loadout second.
8 entries
STATE
Disk persists across sleeps
Chat history, uploaded files, USER.md, MEMORY.md, the FTS5 sessions DB, cron schedules, Python venv, learned skills -- all written to /home/machine. The microVM hibernates between requests; nothing in RAM matters because the disk is the source of truth.
ACCOUNT
One fleet per Clerk identity
Sign in once; your machines follow you across devices and browsers. UserConfig lives in Clerk private metadata -- which provider keys you've attached, which machine is active, which agent is selected. Per-user multi-tenancy from line one.
RUNTIME
Sleep / wake by the second
Dedalus microVMs hibernate when idle and wake on the first prompt -- typically <30s cold-start, <5s after recent activity. Billed by the second. Wake-on-read is wired into chat, artifacts, and the live dashboard.
PROVIDERS
Dedalus live; more hosts shaped
Dedalus Machines is wired end-to-end today. Vercel Sandbox and Fly Machines already exist in the MachineProvider schema and setup UI, with explicit not-supported responses until their provisioners land.
AGENTS
Hermes or OpenClaw
Two agents, same OpenAI-compatible /v1/chat/completions endpoint. Hermes ships memory + cron + MCP-native; OpenClaw ships Anthropic computer-use loop with browser + screenshot. Swap from the dashboard navbar.
TOOLS
23 built-ins + 17 service routes
Terminal, filesystem, web search, browser automation, vision, image generation, code execution, subagent delegation. Cursor is one path; Vercel, Stripe, Supabase, Linear, GitHub, Slack, PostHog, Sentry, Figma, and more sit in the MCP/CLI/skill hierarchy.
KNOWLEDGE
95-skill bundled library
SKILL.md docs auto-loaded by intent: agent-ethos, empirical-verification, production-safety, design-taste, deepsec, torvalds. Drop a folder into knowledge/skills/, click Reload on the dashboard, the VM git-pulls and rsyncs -- no CLI ever.
DELEGATION
Cursor is optional delegation
When configured, Hermes can hand code work to a Cursor SDK subagent through cursor_agent. Without CURSOR_API_KEY, the machine still keeps chat, tools, skills, cron, browser automation, files, and provider control working.
RUNTIME -- LIVE
What you watch on the dashboard.
6 panels . live every 5s
every panel above is rendered from the same /api endpoints the dashboard polls
LOADOUT
Your agent's complete kit.
Mirrors the wiki's tool-hierarchy.mdc: every tool the agent picks from, ranked. Built-in tools fire in one turn. MCP servers get auto-spawned at bootstrap. Service entries cover 17 platforms (Vercel, Stripe, Supabase, Sentry, ...). Task entries rank which skill or tool to use for code review, design review, QA, research, and more.
callable tools
33
skills
95
services
17
task categories
12
built-in tools . 23
view >- Browser6
- Filesystem4
- Memory3
- Search3
- Shell1
- Vision1
- Image1
- Audio1
- Code1
- Delegate1
- Schedule1
services . 17
view >
Vercel
Stripe
Supabase
Clerk
Firebase
Figma
PostHog
Sentry
Datadog
Linear
Slack
Shopify
tasks . 12
view >- Browser automationagent-browser
- Frontend verificationagent-browser diff
- Generative UIjson-render
- Code reviewcode-review
- Design reviewdesign-review
- QA + testingqa
COMPONENTS -- LIVE
The same chrome the dashboard uses.
reticle / sigil
status pill
StatusPill
Click to cycle
4 visible phases. Active phases pulse the dot. Sharp corners, inherits color from the variant.
button
ReticleButton
Three variants, two sizes
Primary uses the single purple. Secondary borrows the border token. Ghost is invisible at rest, surface on hover.
prompt
StarterGrid
One-tap starters
log tail
LogsTail
Polling every 7s
| 12:04:18 | info | POST /v1/chat/completions |
| 12:04:18 | info | loaded 95 skills + 10 mcp tools |
| 12:04:21 | info | stream open . tokens=312 |
| 12:04:21 | warn | cron next: weekly-skill-audit in 3d |
| 12:04:23 | info | stream done . duration=1832ms |
dedalus+nous/openclaw+cursorSKILLS -- 95 BUNDLED
Loaded into ~/.hermes/skills/
all on disk . zero RAG
philosophy
11
How the agent thinks about code -- minimal-fix bias, invariants, taste.
engineering
18
Conventions and refactors -- commits, comments, types, language style.
review
13
Bug-finding and audit discipline -- adversarial review, perf, postmortems.
design
23
Frontend craft -- React, taste systems, component design, image direction.
content
8
Writing -- copy, voice, social, SEO/GEO.
ops
18
Run things, schedule things, browse the internet, query databases.
delegation
4
When and how to spawn another agent (Cursor, sub-agents, skill discovery).
ARCHITECTURE
The machine is the product boundary.
23 nodes, ten rows. The persistent Linux machine in the middle is the product boundary -- everything above provisions and routes to it, everything below runs inside it. Click any node to inspect; drag to rearrange.
machine state
/home/machine persists
Sleep pauses compute. Disk remains the source of truth.
path split
four roots, no overlap
.agent-machines (app), .hermes (runtime), .openclaw, repo checkout.
providers
dedalus + 2 stubs
Dedalus is live. Vercel Sandbox + Fly Machines accept credentials.
loadout
more than cursor
23 built-ins + 17 service MCPs + 95 skills + optional cursor.
FAQ
Common questions about persistent agent machines.
11 answers
- 01
What is Agent Machines?
- Agent Machines is a per-account runtime for persistent agents. Each signed-in user can keep machines with durable Linux filesystems under /home/machine, so chats, working files, artifacts, learned skills, and cron schedules survive sleep and wake cycles.
- 02
How is this different from a regular chatbot?
- A regular chatbot usually stores memory in browser state or a vendor-owned memory layer. Agent Machines persists operational state to a real machine filesystem: chat records, artifacts, USER.md, MEMORY.md, Hermes sessions, cron schedules, skills, and the runtime venv.
- 03
Which agents can I run?
- Hermes and OpenClaw are the two agent runtimes represented in the app. Hermes is the default memory, cron, sessions, and MCP-native runtime. OpenClaw is the computer-use runtime with browser, screenshot, click, shell, and file operations. Both sit behind the same machine/gateway concept.
- 04
Which providers can host the machine?
- Dedalus Machines is wired end-to-end today. The MachineProvider abstraction, setup UI, and user config schema also include Vercel Sandbox and Fly Machines, but those provisioners currently return explicit not-supported responses until their provider implementations land.
- 05
How do I get my own machine today?
- Sign in with Clerk, add provider credentials in /dashboard/setup, pick the agent, provider, spec, and model, then provision the machine record. The browser flow creates the provider machine and stores it in your fleet; the reliable agent bootstrap path is still the matching root CLI deploy command until browser-driven bootstrap lands.
- 06
What tools and skills come pre-installed?
- The public loadout tracks 23 built-in tools, 17 service routes, and 95 SKILL.md files. The surface includes terminal, filesystem, browser automation, web search, vision, image generation, code execution, cron, memory, sessions, Vercel, Stripe, Supabase, Linear, GitHub, Slack, PostHog, Sentry, Clerk, Firebase, Figma, Shopify, ClickHouse, Datadog, AWS, Cloudflare, and model providers.
- 07
Is Cursor required?
- No. Cursor is optional delegation for code edits through cursor-bridge and @cursor/sdk. Without CURSOR_API_KEY, the rest of the machine still runs: chat, files, browser automation, skills, cron, memory, dashboard polling, artifacts, and provider lifecycle controls.
- 08
What are ~/.agent-machines, ~/.hermes, and /home/machine/hermes-machines?
- ~/.agent-machines is Agent Machines product data such as chats and artifacts. ~/.hermes is Hermes runtime state such as skills, crons, sessions, logs, MEMORY.md, USER.md, and config. /home/machine/hermes-machines is the git checkout used by reload-from-git.sh. They are separate on purpose.
- 09
What inference providers are supported?
- Dedalus is the default OpenAI-compatible inference endpoint. Hermes is configured through model.base_url and model.default, so the CLI can point DEDALUS_CHAT_BASE_URL at another compatible /v1 endpoint when needed. The dashboard stores a model slug per machine.
- 10
What happens when a machine sleeps?
- On supported providers, sleep pauses compute while preserving the persistent volume. The next wake resumes from disk: app artifacts, agent runtime state, skills, cron schedules, sessions, and the venv remain available.
- 11
Where does my data live?
- Provider credentials and gateway bearers live in Clerk private metadata. Machine state lives on the provider machine under /home/machine, with Agent Machines product data under ~/.agent-machines and Hermes runtime data under ~/.hermes. The public client only sees redacted provider and machine status.