Free preview

A sandbox for AI agents. On your machine.

A bounded, inspectable space to run the coding agents you already use — several of them at once, wired into a flow, pointed at repos that never have to leave hardware you control.

Agent CLIs
Claude Code, Codex, Copilot, Cursor, Antigravity
Tooling
MCP-native, one gateway per workspace
Where it runs
Your own machine, or hosted by us
Free tier
Unmetered — your machine, your key

Sandbox · your machine

  1. coder writes the change
  2. reviewer reads it back, hands findings over
  3. qa exercises it, records a verdict

repo ~/work/api tools via MCP data stays local

Why this exists

Two frictions, one box

If you run an agent CLI every day, you have probably hit at least one of these. They are the whole reason AW is shaped the way it is.

Friction 01

One agent at a time is not a pipeline

Writing, reviewing and verifying a change are different jobs with different instructions. Stuffing all three into one prompt gets you an agent that marks its own homework.

→ AW runs them as separate agents and routes the handoff between them.

Friction 02

Your repo, someone else's black box

Pointing an autonomous agent at real source means deciding where that source sits while it works, and what you can see of what the agent did to it.

→ AW can run on a machine you own, reached over a tunnel you link yourself.

Orchestration

A flow, not one giant prompt

Agents Flow is a graph: you draw which agent hands off to which, and each hop is a real, separately-instructed agent with its own contract. A coder opens the change, a reviewer reads it back, QA exercises it and records the verdict.

  • Explicit handoffsEvery hop ends with a decision — hand off, report back, or declare the task done.
  • Pick the model per roleRun the reviewer on a different model from the coder when that is the right trade.
  • Readable afterwardsRuns, handoffs and verdicts are recorded, so you can see what each agent actually did.

Flow · handoff

illustrative

coder → opens the change · hands off reviewer → reads it back · returns findings coder → applies the fixes · hands off qa → exercises it · records a verdict flow done

On your machine

Bring your own box

Link a machine you already own and the workspace runs there: your repos, your filesystem, your network. The console reaches it over a tunnel you established, so the code an agent touches stays where you put it. Prefer not to run anything yourself? Host it with us instead — same product, same console.

  • Your filesystem, mappedPoint the workspace at any directory on that machine, git checkout or not.
  • Your CLI loginsThe agent CLIs authenticate from your workspace, against your own accounts.
  • One console either wayHosted and bring-your-own-machine workspaces are managed from the same place.
The console reaches a workspace running on your own machine over a tunnel you link. YOUR MACHINE workspace ~/work/api agent CLIs tunnel CONSOLE flows runs billing
Your code stays on your side of the tunnel

What's actually inside

Checkable, not claimed

No logos we have not earned and no quotes we made up. Four things that are simply true of the product today.

01

MCP-native throughout

Tools reach agents over the Model Context Protocol, aggregated behind one gateway per workspace rather than a config file per CLI.

02

Drives the CLIs you use

It installs and runs the real agent CLIs, each authenticated against your own accounts.

  • claude
  • codex
  • copilot
  • cursor-agent
  • agy

03

Multi-agent by design

Agents, flows and scheduled tasks are first-class: a flow is a graph of agent-to-agent handoffs, not a prompt chain stapled together.

04

Extensible in the open

Apps install into a workspace and contribute their own windows, tools, CLIs and skills — the same mechanism we build on ourselves.

Deliberately absent: customer logos, testimonials and any user, revenue or uptime figure. AW is in free preview — when there are real numbers and real quotes, they will show up here with names attached.

Pricing

Free where you host it. Metered where we do.

On your own machine, usage runs free because we don't bear the cost. Hosted by us, you pay only for what you use.

Free

Confirmed

$0 /month

Unmetered agent runs Unmetered images

Runs on your own machine with your own LLM key, so we pay no compute and count nothing. The only real gate is tunnel bandwidth — PLACEHOLDER GB/month, not set yet.

  • Runs on your own machine
  • Multi-agent flows
  • 1 workspace, up to 5 members
  • No card to start
Start free

Paid

Placeholder

$— /month — PLACEHOLDER

Bundled monthly quota + pay-as-you-go overage

The shape is decided; the numbers are not. Quota size and overage rate are PLACEHOLDER until a run has a published cost ceiling. New accounts also get a one-time, 30-day hosted trial (up to 200 agent runs + 50 images, once per account) — described here, not yet available to start.

  • A larger bundled run quota (size: PLACEHOLDER)
  • Per-run overage past the quota (rate: PLACEHOLDER)
  • One-time 30-day hosted trial — up to 200 runs + 50 images (not yet available)
  • Spend surfaced to the agent as it runs
  • Everything in Free
Start on Free

Teams

Placeholder

$— /month — PLACEHOLDER

Pooled quota + pay-as-you-go overage

Same metering, shared across the team instead of per person. Every figure on this tier is PLACEHOLDER.

  • One pooled quota for the team (size: PLACEHOLDER)
  • Per-run overage past the pool (rate: PLACEHOLDER)
  • Shared workspaces and flows
  • Usage visible per member
Start on Free

PLACEHOLDER notice, stated plainly: the free-tier quotas above are real and confirmed. No paid price has been set. Agent runs vary in cost by more than an order of magnitude, so publishing a paid number before a run has a defined ceiling would be a figure we invented rather than one we measured. The structure — a bundled quota plus metered overage — is what we are building to.

FAQ

The obvious questions

No — different customer. Those sell sandbox infrastructure to people building agent products, as an API with no end-user application. AW is the sandbox you run your own agents in: a console, a workspace, a flow editor, and the option of running the whole thing on your own machine.

Claude Code, OpenAI Codex, GitHub Copilot CLI, Cursor agent and Google Antigravity CLI are installed into the workspace and driven from it. Each still needs its own login, run once from a workspace terminal, against your accounts.

You link a machine you own, and the workspace — repos, filesystem, agent CLIs, installed apps — runs there. The console connects to it over that tunnel to dispatch and observe runs. If you would rather not host anything, choose a hosted workspace instead; the product is the same.

One dispatch of one agent — a coder run, a review run, a QA run. A three-hop flow therefore spends three. The precise per-run ceiling is still being set, and it is the same open question that keeps the paid prices above marked PLACEHOLDER.

Yes — Free is capped at 1 workspace and up to 5 members. What it does not cap is usage: Free runs on your own machine with your own LLM key, so we pay no compute and don't count your agent runs or images.

If one agent in your editor already does what you need, AW is more machinery than the job calls for. It starts earning its keep when you want several agents with different instructions handing work to each other, or when where the code runs is a decision you want to make yourself.

Open the box and look inside

Free preview: your own machine, your own key, runs not metered by us. No card to start — 1 workspace, up to 5 members.