llms.txt
Compare with other sites' llms.txt files →# Agent Machines
> Open-source building blocks for agent harnesses.
Agent Machines connects runtimes, tools, configuration, managed state, and
remote compute in a working implementation developers can inspect, adopt,
tune, and extend. It includes a real browser agent CLI and a dashboard for
operating multiple configured agents.
“shadcn for agent harnesses” describes the source-first adoption model. It does
not mean shadcn registry compatibility, a component-install CLI, independently
packaged components, or a public Worker marketplace.
## Canonical pages
- [Website](https://agent-machines.dev)
- [Building blocks](https://agent-machines.dev/components)
- [Documentation](https://agent-machines.dev/docs)
- [Source and quick start](https://github.com/Kevin-Liu-01/Agent-Machines)
- [Product messaging and claim ledger](https://github.com/Kevin-Liu-01/Agent-Machines/blob/main/docs/PRODUCT-MESSAGING.md)
- [Architecture whitepaper](https://github.com/Kevin-Liu-01/Agent-Machines/blob/main/docs/WHITEPAPER.md)
- [Roadmap](https://github.com/Kevin-Liu-01/Agent-Machines/blob/main/docs/ROADMAP.md)
- [Release checks and dated evidence](https://github.com/Kevin-Liu-01/Agent-Machines/blob/main/docs/LAUNCH.md)
## Implemented building blocks
- Runtime adapters: Claude Code, Codex, Hermes, OpenClaw.
- Compute adapters: Daytona, E2B, Sprites, Vercel Sandbox.
- Runtime bootstrap and supported model/environment configuration.
- Browser Agent Console: the real remote agent CLI, using worker-owned tmux
sessions with native terminal transports or a portable fallback.
- Hosted Memory documents, pasted import, instruction editing, and selected
abilities; runtime-specific document installation.
- Multiple-agent dashboard with recipes, lifecycle operations, terminals, files,
logs, sessions, loadouts, schedules, and usage.
- Direct provider SDK, hosted client, and a declarative lifecycle kernel.
## Catalog facts
The bundled catalog includes 161 skills and 39 MCP servers. Counts come from
the same committed registries used by the dashboard:
| Catalog | Entries |
|---|---|
| SKILL.md skills | 161 |
| Service lanes | 27 |
| MCP servers | 39 |
| CLI catalog entries | 24 |
These are available resources, not proof that every tool is installed,
authenticated, compatible, or production-tested. Service lanes are configured
interface rankings; CLI entries are discoverable tools, not installed binaries.
## Package and source boundaries
The MIT-licensed package is named `agent-machines`.
- `AgentMachines` from `agent-machines`: authenticated client for a configured
deployment. `create()` waits for launch completion; `run()` waits for a result.
The account needs its own sandbox and compatible model credentials.
- `createMux` from `agent-machines/mux`: direct provider access with your own
credentials; a hosted Agent Machines account is not required.
- `AgentMachinesControlPlane`, stores, and `MuxWorkerRuntimeDriver` from
`agent-machines/control-plane`: desired state, operations, and reconciliation.
The application must run the reconciler.
- The default public mux driver does not install hosted Memory/loadouts or apply
hosted role/gateway/environment-profile fields merely because the type includes
them.
- Browser terminal, dashboard, hosted bootstrap, and Memory implementations are
source modules in the web app, not independent component packages.
- `src/mux/loadout.ts` is source-only and is not integrated into the default
router or compiled public package. It is not a supported package import.
The hosted client uses `AGENT_MACHINES_URL` and
`AGENT_MACHINES_API_KEY`. The current direct production API origin is
`https://www.agent-machines.dev`; clients should not rely on cross-origin
redirects to carry authorization. Timing out stops the local wait, not already
accepted server work. The current hosted launch always owns bootstrap;
`bootstrap: false` only affects the separate request used by legacy servers.
### Hosted client configuration
This example constructs a client only. It sends no request and does not
provision compute. Keep the API key in the server-side process environment.
```ts
import { AgentMachines } from "agent-machines";
const am = new AgentMachines({
baseUrl: "https://www.agent-machines.dev",
apiKey: process.env.AGENT_MACHINES_API_KEY,
});
```
- [Public exports](https://github.com/Kevin-Liu-01/Agent-Machines/blob/main/src/index.ts)
- [Direct mux reference](https://github.com/Kevin-Liu-01/Agent-Machines/blob/main/docs/MUX.md)
- [Lifecycle boundaries](https://github.com/Kevin-Liu-01/Agent-Machines/blob/main/docs/CONTROL-PLANE-V2.md)
## Compatibility and state limits
Four runtime adapters and four provider adapters are implemented; this is not a
claim that every combination has passed a current live test.
Hosted routing uses native Anthropic for Claude Code and native OpenAI for
Codex, with supported gateway paths for Hermes/OpenClaw. The direct mux has a
different mapping: Claude Code, Codex, and OpenClaw accept declared compatible
gateways; its Hermes adapter is native-only. Arbitrary endpoints are not
universally compatible.
Daytona stop/start retains files, not processes or RAM. Processes restart after
stopping. E2B supports manual
pause/resume and uses template-defined resource allocation. Sprites uses
provider-managed idle suspension without a manual Sleep operation. Vercel
Sandbox restores filesystem snapshots, not running processes or RAM. Runtime
capacity, available credentials, and transport support must be checked.
Daytona's sandbox home is `/home/daytona`. Archived measurements retain their
original providers and dates; they are not Daytona measurements. Daytona has
no verified compute rate or inherited timing in the current router price model;
it does not borrow a retired provider's rate or benchmark.
Migration transfers allowlisted durable state and restarts processes. Placement
failover does not automatically replay interrupted agent runs. Hosted routing
features and direct SDK routing features are not identical.
## Terminology and non-claims
A **Worker** is the dashboard/API name for a configured agent workspace and
its lifecycle intent. Lasting configuration and managed files make a harness
reusable; this is not a promise of autonomous labor or unlimited uptime.
A **recipe** is a preconfigured starting point. A **loadout** is selected
abilities, not proof of installation or connection. The **registry** discovers
tools, skills, integrations, and supported imports; it is not a public
marketplace for complete harnesses.
Memory export produces Markdown instructions plus ability labels, not a full
executable harness/configuration archive. Provider and model accounts and
charges remain separate.
Natural-language harness creation, full harness publishing/sharing, a public
Worker marketplace, authority-scoped autonomous delegation, and unified provider
billing are directions, not shipped features.
## Evidence
Local tests and package checks prove their checked contracts, not live provider
behavior. The September 9 reports include a new-account Codex-on-Daytona task
with artifact read-back. The August 5 strict matrix is historical and includes
a retired provider; it is not Daytona validation. Preserve dates, runtime,
provider, resource, output, and cleanup scope when citing a live result.
- [Dated provider measurements](https://github.com/Kevin-Liu-01/Agent-Machines/blob/main/docs/MUX-RESULTS.md)
- [Release procedure and reports](https://github.com/Kevin-Liu-01/Agent-Machines/blob/main/docs/LAUNCH.md)