CLI examples

Each sequence runs as written after npm install -g @tangle-network/sandbox-cli and tangle auth login (or TANGLE_API_KEY in the environment). Commands take a sandbox id or its name. Each sequence that creates a sandbox ends by deleting it. Every flag is listed in the CLI reference.

Fix a test and keep the patch

tangle sandbox create --name fixer --backend claude-code \
  --git-url https://github.com/tangle-network/sandbox-demo
tangle agent prompt fixer "The test tagged null-guard fails. Fix the cause." --stream
tangle exec fixer -- npm test
tangle exec fixer -- git diff > fix.patch
tangle sandbox delete fixer --force

--stream prints the agent’s work as it happens. tangle exec exits with the command’s exit code, so a script can stop when the tests still fail. Guide: Run a coding agent.

Pair with a coding agent in its own terminal UI

tangle sandbox create --name pair --backend claude-code \
  --git-url https://github.com/tangle-network/sandbox-demo
tangle agent prompt pair "Find the failing test and fix it." --interactive
tangle agent prompt pair "Review the last change." --interactive --harness codex
tangle sandbox delete pair --force

Your keys go to the agent; Ctrl-] detaches and leaves the session running. The CLI prints a dashboard link where you can watch and drive the same session in a browser. Guide: Interactive mode.

Continue one conversation over several turns

tangle sandbox create --name planner --backend codex \
  --git-url https://github.com/tangle-network/sandbox-demo
tangle agent prompt planner "Read the codebase and plan input validation for the CLI." --session plan
tangle agent prompt planner "Implement step 1 of your plan." --session plan --stream
tangle sandbox stop planner
tangle agent prompt planner "Implement step 2." --session plan --stream
tangle sandbox delete planner --force

The same --session keeps the conversation. agent prompt resumes the stopped sandbox before the next turn, and the files and the conversation are still there. Guide: Sessions.

Try a change and roll it back

tangle sandbox create --name trial --backend codex \
  --git-url https://github.com/tangle-network/sandbox-demo
tangle snapshot create trial --tags clean
tangle agent prompt trial "Replace the date library with the built-in Intl API." --stream
tangle exec trial -- npm test
tangle snapshot list trial
tangle snapshot revert trial <snapshot-id>
tangle sandbox delete trial --force

snapshot create prints the snapshot id. snapshot revert puts the whole disk back as it was. Guide: Snapshots.

Triage a list of tickets in one batch

tangle batch run --backend triage=claude-code \
  --task t1="Classify as frontend, backend or email: export button does nothing in Safari" \
  --task t2="Classify as frontend, backend or email: invoice shows the wrong currency" \
  --task t3="Classify as frontend, backend or email: password reset email arrives twice" \
  --stream

The platform creates the sandboxes, runs every task, retries failures and cleans up. Add --idempotency-key <key> to rejoin the run after a disconnect. Guide: Batches and hosted runs.

Run several agents under a budget

tangle fleet run "Write a prime checker in Python with tests, run the tests and report the result." \
  --agents 3 --placement dedicated --max-spend-usd 1

To give each agent its own task and coding agent, such as Claude Code, Codex or OpenCode, write the agents to a file:

[
  { "agentId": "py", "harness": "claude-code", "task": "Write a prime checker in Python with tests and run them." },
  { "agentId": "js", "harness": "codex", "task": "Write a prime checker in Node with node:test and run it." },
  { "agentId": "go", "harness": "opencode", "task": "Write a prime checker in Go with tests and run them." }
]

tangle fleet run --agents agents.json --placement dedicated --max-spend-usd 1

Prints one line per agent with its coding agent, model, status, time, cost and answer, then deletes the fleet. Guide: Fleets.

Supervise workers under one budget

tangle supervise "Write a small URL shortener in Node with tests, and make the tests pass" \
  --harness claude-code --max-tokens 200000 --max-iterations 50 --max-workers 3

The supervisor spawns workers, each in its own sandbox, and stops when the token or iteration budget runs out. It prints the result, tokens and cost. Guide: agent-runtime.

Let an agent read your GitHub issues

tangle hub connect github
tangle hub connections
tangle sandbox create --name triager --backend claude-code \
  --git-url https://github.com/tangle-network/sandbox-demo
tangle agent prompt triager "List my open GitHub issues and say which one this repository could fix." \
  --connection github:* --stream
tangle sandbox delete triager --force

hub connect opens a browser to sign in to the provider. --connection github:* grants the turn every action of that connection. Read actions run; every other action follows its permission (allow, ask or deny). Guide: Connections.

Script with JSON output

id=$(tangle sandbox create --backend claude-code --json | jq -r .id)
tangle agent prompt "$id" "Say hello in one word." --json | jq .
tangle sandbox delete "$id" --force

With --json, stdout carries only the JSON document, and an error prints a JSON envelope with its code. See the CLI reference for the commands that always print text, and Errors for every code.

Examples Coding agent examples