Run harness adapters inside isolated sandboxes with declarative provisioning and durable run takeover.
npx skills add https://github.com/TanStack/ai --skill ai-sandbox
Please help me install this Agent Skill. Skill: ai-sandbox Source: https://github.com/TanStack/ai/tree/main/packages/ai-sandbox/skills/ai-sandbox Command: npx skills add https://github.com/TanStack/ai --skill ai-sandbox
Running agent CLIs like Claude Code, Codex, or OpenCode directly on the host is unsafe and hard to reproduce. This Skill runs harness adapters inside isolated sandboxes (local process, Docker, Daytona, Cloudflare) with declarative workspace provisioning, secret handling, snapshots, and durable run journals so long-running agent runs survive disconnects and host failures.
Use the ai-sandbox skill to run Claude Code inside a Docker sandbox with a git workspace, setup steps, and snapshot-after-setup enabled.
High-intent search queries and answers about installing and using this skill.
Define a sandbox with defineSandbox using dockerSandbox({ image }) as the provider and a defineWorkspace with your git source and setup commands, then pass it to chat() via the withSandbox middleware. The adapter runs the agent CLI inside the container and streams events back.
When the provider supports snapshots, bootstrap takes one automatically after setup completes, and subsequent runs resume from it. You can set snapshotMaxAge as a TTL, and portable snapshots via withPersistence let you save named checkpoints, fork from them, and read artifacts.
Three providers are documented: localProcessSandbox for host execution during development, dockerSandbox for isolated containers with snapshots and fork, and daytonaSandbox for Daytona cloud sandboxes. All implement the same SandboxHandle interface with fs, git, process, and ports.
When sandbox durability is wired, every run must receive an explicit unique runId because the journal path is derived from it and no generated id can be recomputed by a successor host. Pass a unique runId per run; reusing one appends to the previous run's journal and corrupts replay.
Yes, when you pass both runs and durability to withSandbox, the agent's NDJSON output is journaled inside the sandbox and sandboxRunDriver enables takeover by a new host. Passing only one of the two options silently disables durability entirely.
createSecrets stores values in a non-enumerable symbol-keyed registry so they are never logged, snapshotted, or written to the sandbox store. Use secret refs in gitSkill for private repos and bearer() for MCP authorization headers, resolved only at bootstrap time.
namuanProvision and manage Docker sandboxes for AI coding agents.
anderskevBuild streaming chat interfaces with tool calls using the Vercel AI SDK.
dadbodgeoffManage distributed locks with atomic acquisition and TTL expiration.
spinlockdevelopmentRun agents inside an isolated Docker sandbox with read-only host mounts.
