AGENT MACHINES

stateful microVMper-account fleet

A persistent machine
for your agent.

One stateful microVM per account. Boot in 30 seconds, sleep on idle, wake on the first prompt.

  • State on disk, not in RAM

    Chats, files, USER.md, MEMORY.md, FTS5 sessions, cron, venv -- everything lives at /home/machine and survives every sleep.

  • Per-account fleet

    Sign in once with Clerk; the same machine wakes on every device. Provider keys + agent choice persist in private metadata.

  • Bring any agent + tool

    Hermes or OpenClaw, any provider key (Dedalus / Sandbox / Fly), any of 95 skills + 17 MCP services -- swap any time.

GitHubDocsDocs

->Each cell on the right is one day this machine was awake. Hover to peek, click to pin. Nothing lives in RAM that it can't rebuild from /home/machine on restart.

ACTIVITY -- 6 MONTHS

115 active days

hover or click a cell

nov
dec
jan
feb
mar
apr
may
lessmore

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

Dedalus Machines

runtime layer

->

Hermes / OpenClaw

agent layer

Hermes ->OpenClaw ->

Cursor SDK

codework layer

->

Vercel

web + ai gateway layer

->

Clerk

auth + fleet metadata layer

->

Cloudflare

public tunnel layer

->

Anthropic

model provider layer

->

OpenAI

model provider layer

->

CAPABILITIES

State persists. Account is the key.

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.

/home/machineMEMORY.mdFTS5

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.

Clerk private metadataMachineRef[]activeMachineId

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.

microVMwake-on-readsecond-billing

PROVIDERS

Bring any key

Dedalus is the default runtime, but the MachineProvider interface accepts Vercel Sandbox and Fly Machines too -- each user attaches their own credentials. Same shape, same agent, same persistence.

dedalusvercel-sandboxfly

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.

/v1/chat/completionsswap any timesame machine

TOOLS

23 built-in + 17 MCP services

Terminal, filesystem, web search, browser automation (Playwright), vision, image generation, code execution, subagent delegation. Cursor, Stripe, Supabase, Linear, Vercel, GitHub, Slack, PostHog, Sentry, ... -- mounted as MCP servers and selectable per machine.

terminalbrowser_*MCPsubagent

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.

agent-ethosdeepsecgit pullReload button

DELEGATION

Spawns Cursor coding agents

When the agent needs to write real code, it hands off to a Cursor SDK subagent via the cursor_agent MCP tool. Same model the Cursor IDE runs, full file access, the machine's selected skills injected as .cursor/rules.

cursor_agent@cursor/sdk.cursor/rules

RUNTIME -- LIVE

What you watch on the dashboard.

disk/home/machine . persists
skillsvenvchatsartifactsmemorycron
2.1 GiB used . 7.9 GiB free
latencylast 32 chat completions
p50 412ms . p95 1.3s
awakelast 24h . second-billed
00:0006:0012:0018:00now
3 wakes . 19m awake
tokensthis conversation
systemuserassistant
prompt 1.2k . completion 3.1k
loadedskills active in context
agent-ethosempirical-verificationtaste-outputreticle-design-systemvercel-react-best-practicescounterfactual
6 of 95 . by intent match
next cronweekly-skill-audit
3d 04huntil next fire
fires in 3d 04:00 utc

every panel above is rendered from the same /api endpoints the dashboard polls

LOADOUT

Your agent's complete kit.

See full loadout

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
each ranksMCP>CLI>skills

tasks . 12

view >
  • Browser automationagent-browser
  • Frontend verificationagent-browser diff
  • Generative UIjson-render
  • Code reviewcode-review
  • Design reviewdesign-review
  • QA + testingqa
skill . design23
skill . ops18
skill . engineering18
skill . review13
skill . philosophy11
skill . content8
skill . delegation4
callable byhermes/openclaw

COMPONENTS -- LIVE

The same chrome the dashboard uses.

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

12:04:18infoPOST /v1/chat/completions
12:04:18infoloaded 95 skills + 10 mcp tools
12:04:21infostream open . tokens=312
12:04:21warncron next: weekly-skill-audit in 3d
12:04:23infostream done . duration=1832ms
built withdedalus+nous/openclaw+cursor

SKILLS -- 95 BUNDLED

Loaded into ~/.hermes/skills/

philosophy

11

How the agent thinks about code -- minimal-fix bias, invariants, taste.

agent-ethoscounterfactualempirical-verificationfail-closed-case-matchinginvariant-first-codinginvariant-first-testing+5

engineering

18

Conventions and refactors -- commits, comments, types, language style.

column-aligned-fieldscommentcommitfix-typesgenerate-interfacegit-workflow+12

review

13

Bug-finding and audit discipline -- adversarial review, perf, postmortems.

bugscode-reviewcross-modal-reviewdaily-bugfix-checkdeepsecexemplar-audit+7

design

23

Frontend craft -- React, taste systems, component design, image direction.

design-engineeringdesign-reviewfont-switcher-dev-toolfrontend-designfrontend-design-tasteloading-screens+17

content

8

Writing -- copy, voice, social, SEO/GEO.

app-onboarding-questionnairecontent-strategycopywritingseo-auditseo-geo-optimizationsocial-content+2

ops

18

Run things, schedule things, browse the internet, query databases.

agent-activity-logagent-browseragent-reachautomation-cronci-cd-best-practicescomputer-use+12

delegation

4

When and how to spawn another agent (Cursor, sub-agents, skill discovery).

cursor-codingfind-skillsskill-auditorskill-creator

ARCHITECTURE

One key, two endpoints, full agent.

The same DEDALUS_API_KEY provisions the machine and authenticates inference. Dedalus owns the runtime, Nous ships the agent, Cursor handles codework.