A coding agent on your repository, with a spending cap

Why it pays

• A clean machine every time No leftovers from the last task, and nothing touches your laptop.
• Spending that stops itself A per-prompt budget ends a runaway turn, at the proxy, before the bill grows.
• Your keys stay outside The agent sees a placeholder; the credentials proxy adds the real key on the way out.

Step by step

1. Connect your model key once

   Store the key with HareBox and route it through the proxy, so it never enters a sandbox.

   CLI

   harebox ai key set default anthropic-api-key --value "$ANTHROPIC_API_KEY"
   harebox proxy rule add --secret-ref anthropic-api-key --host api.anthropic.com --header x-api-key --template "{secret}"

2. Start a sandbox for the issue

   Use an environment that holds the repo and setup, and label the sandbox with the issue.

   CLI

   ID=$(harebox new --name issue-482 --environment web-app --label issue=482 | jq -r .id)

   Python

   box = hb.create(name="issue-482", environment="web-app", labels={"issue": "482"})

   TypeScript

   const box = await hb.create({ name: "issue-482", environment: "web-app", labels: { issue: "482" } });

3. Hand the agent the task, with a cap

   Pick the harness and a dollar budget. Events stream as the agent works.

   CLI

   harebox prompt "$ID" "Fix issue 482, add a test, and open a pull request" --harness claude --budget 3

   Python

   for ev in box.prompt("Fix issue 482, add a test, and open a pull request",
                        harness="claude", budget_usd=3):
       print(ev["type"])

   TypeScript

   for await (const ev of box.prompt("Fix issue 482, add a test, and open a pull request",
     { harness: "claude", budget_usd: 3 })) console.log(ev.type);

4. Steer it, if it drifts

   Add guidance mid-turn without starting over, or interrupt just this conversation.

   CLI

   harebox steer "$ID" --message "Keep the public API unchanged"
   harebox interrupt "$ID"

   Python

   hb.steer_agent(box.id, body={"message": "Keep the public API unchanged"})

   TypeScript

   await hb.steerAgent(box.id, { body: { message: "Keep the public API unchanged" } });

5. Check the cost and stop

   Read the exact bill for the task, then stop the machine to end it.

   CLI

   harebox usage "$ID"
   harebox stop "$ID"

   Python

   print(hb.get_sandbox_usage(box.id)["total_usd"])
   box.stop()

   TypeScript

   console.log((await hb.getSandboxUsage(box.id)).total_usd);
   await hb.stopSandbox(box.id);

Tips

• Conversations live on the sandbox's disk, so a stopped sandbox resumes with the agent's memory.
• Use --new for a fresh conversation in the same machine, and run several in parallel.
• Set a project budget as well as per-prompt caps, so a busy week has a ceiling too.