Free modelsGemma 4 31B + Gemini 3.6 Flash on every pad

Give coding agents everything they need to build software.

Reachpad gives Claude Code, Codex, OpenCode and other coding agents persistent cloud development environments where they can build, run and test software independently.

brew install reachpad/tap/reachpad

A formula, so brew upgrade keeps it current.

Start trialRead the quickstart
Coding agents reach a reachpad workspace through the CLI, the MCP server or the REST API. The workspace holds your repository on a branch, its dependencies, its databases and services, and the dev server on its ports. Code leaves through your own git remote.claude-codecodexopencodeCLI · MCP · RESTworkspacerepodepsservices:3000pushyour git remote

Start the work. Come back to a result.

Every one of these keeps running after you close the terminal. Nothing needs your attention until there is something to read.

A repository you have never built

setup would take your afternoon

A branch that builds, and the test run that proves it

A bug that will not reproduce

your machine does not have the database

The failing test, and the commit that makes it pass

A refactor that runs for hours

your terminal would have to stay open

Every call site changed, in one diff you can read

A backlog worked overnight

you are asleep

One branch per ticket, each with the test run behind it

Five agents on one ticket. You choose what lands.

One known-good checkout forks into five workspaces, each running an agent on the same ticket at the same time. Three come back with a passing suite and two do not. You review the three.add-indexbatch-loadercache-readdenormalizelazy-join
~/storefront

$ reachpad fork ws_4f1a --count 5 | \

xargs -P5 -I@ reachpad run @ -- \

sh -c 'claude -p "$TICKET" && npm test'

One id per line, so it pipes. Each run exits with the suite's own code, so the shell keeps score and you read the diffs.

Every fork

starts at the parent's last save, disk and memory both, so no agent reinstalls a dependency or waits for the database to come up

Five forks

hold five workspace slots while they run. Archive the ones you do not keep and the slots come back

Use the agents and workflows you already have.

Reachpad does not replace your agent. It gives it somewhere to work.

MCPhttps://reachpad.dev/mcp

Claude and ChatGPT

One connector, and the model creates and forks environments as tools.

Connect a model
CLIbrew install reachpad

Your terminal

Create, run and fork environments from your own shell.

Install the CLI
RESTPOST /v1/environments

Your own agent

The same environments over HTTP, from your own agent.

Read the API docs

What you actually use

You pay for compute that is doing something. Set your rate of work and see how far the free grant goes, and how many hours a month you would be billed for once it does not.

4h
5d
2×

A paused workspace costs nothing, so only active compute is counted here.

Free grant lasts2.1 daysof work at this rate
Then you bill173 h/monthof active compute
Idle time billed0 han always-on machine would bill 1,460 h

The numbers, in full

What the fleet enforces, not what it aspires to. Every figure here is read from the code that enforces it.

Free trial

Workspaces per account10(on the free trial)
Running at once5(the rest stay paused)
Compute grant1,000 credits(16 hours 40 minutes of active work)
Cost of a paused workspace$0(state is kept, the meter stops)
Card required to startNo

Beyond the grant

Billing unit1 credit(one workspace, one minute of active compute)
Billed while pausedNothing(state is kept)
Billed for idle timeNothing
Monthly minimumNone
Seat countNot billed(compute is)
Agents running side by sideSeveral(one tab each)

Check it yourself

Command-line clientPublic source(read the client you install)
Model Context Protocol serverPublished(@reachpad/mcp on npm)

Full pricingHow reachpad bills a credit