Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

sandboard
open source / self-hosted Get started

an open-source platform for coding agents

Run coding agents
in OpenShell sandboxes.

Sandboard coordinates coding agents. OpenShell provides the isolated execution environment. You create a Project, dispatch its Tasks, and review the resulting changes or pull request on GitHub.

Read the quickstart See how it works
run it locally git clone https://github.com/sandboard-app/sandboard.git
sandboard / operator view WORKFLOW
SANDBOXED AGENT WORKOpenShell · policy · provider
Backlog2
PLAN-01Shape the next change
TASK-04Waiting on its dependency
Running1
TASK-03Agent is working
Needs You1
TASK-02A decision is waiting
Review1
TASK-01Pull request is ready
Done12
MERGEDMerged on GitHub
board / sandbox / githubreview on github

what sandboard is

The board for
coding agents.

Sandboard is the part you operate: the control plane for repository work. It stores Projects and Tasks, provides the operator UI and MCP endpoint, and moves each Task from backlog to review.

OpenShell is the isolated execution runtime. Sandboard uses its gateway to create the sandbox, select the image, apply the network policy, inject provider credentials at runtime, and start the agent.

separate responsibilitiesSandboard controls lifecycle; OpenShell controls execution

how a Task moves

Sandboard starts the run.
OpenShell isolates it.

The board stores state and operator actions. Sandboard uses the OpenShell gateway to create the isolated environment, watch the run, and collect the agent’s result.

sandboard / live tracecard 01
01 / createCreate a Project and Task. 02 / provisionOpenShell creates the sandbox. 03 / observeSandboard tracks the run. 04 / reviewReview the pull request.
sandboard tracks output●
01 / create

Create a Project.
Describe the Task.

Create a Project, point it at a repository, and explain what you want. Sandboard creates an Initial plan that you can edit before it becomes implementation work.

Understand Projects and Tasks
02 / provision

OpenShell
creates the sandbox.

When you dispatch a Task, Sandboard asks the OpenShell gateway to provision a sandbox with the selected image and network policy, make the configured provider available, and start the agent.

See how sandboxes work
03 / observe

Sandboard
tracks the run.

Sandboard observes the agent’s output, keeps the card state current, and collects a plan, report, escalation, or split artifact. The worker cannot call the board directly.

Learn the daily workflow
04 / review

Review the
pull request.

Completed work arrives with its pull request and evidence. Sandboard can surface the change, but you merge it on GitHub.

Read the invariants

who owns what

Sandboard owns the work.
OpenShell runs the agent.

01

Sandboard

Stores Projects and Tasks, serves the UI and MCP endpoint, and runs the supervisor that owns lifecycle transitions.

02

OpenShell

Creates the sandbox and applies the image, network policy, and provider credential binding.

03

Agent

Works inside the sandbox with repository access and no network path back to Sandboard.

04

GitHub

Receives the pull request. You review and merge it; Sandboard does not merge it for you.

the complete setup

Run Sandboard with OpenShell.
Then dispatch agent work.

Start the board locally, connect an OpenShell gateway, choose a sandbox spec, and attach a provider. Without those pieces, Sandboard can show the board but it cannot run an agent.

Quickstart Read the tour First agent
local / first run01—03
git clone https://github.com/sandboard-app/sandboard.gitcd sandboardcargo run