Cap a Claude Code session in tokens or dollars. When spend crosses the line — your subagents counted — the run halts cleanly. It auto-arms $50 the moment you start a /goal, and trips when an agent gets stuck in a loop.

Native hook. No proxy, no API key, nothing leaves your machine, fully within Anthropic's Terms of Service.

Claude Code now runs autonomously for long stretches — /goal loops until a condition is met, agents retry, subagents fan out. That's powerful, and it's exactly how a session quietly burns through your budget: a stuck loop or a goal that never converges can run up your tokens (or, on API billing, real dollars) while you're away from the keyboard.

Claude Code shows usage after the fact (/usage) but it won't stop. burnstop is the missing circuit breaker: set a cap, and when the session crosses it, the run halts — even mid-/goal.

As a plugin (recommended):

/plugin marketplace add phuryn/burnstop
/plugin install burnstop

Or locally — and the recommended path on Windows (where python3 is awkward and Claude Code runs hooks through bash):

git clone https://github.com/phuryn/burnstop
cd burnstop
python cli.py install      # registers the hooks + /budget commands with your own python

Use one path or the other, not both. After installing, restart Claude Code (or open a new session) — commands hot-reload into a running session, but hooks load at session start.

Just start a goal — it auto-arms. /goal ... caps the goal at $50 automatically (change with /budget-default $100; an explicit budget is never overridden). A runaway goal halts at the cap.

Set a budget yourself — type /budget to discover them all:

/budget-arm $1          cap from now ($ or tokens), incl. subagents
/budget-status          spend vs cap
/budget-reset           fresh allowance, keep the cap
/budget-disarm          remove the cap
/budget-default $100    set the budget auto-armed on /goal

Budgets accept dollars ($1, 1$, $0.10) or tokens (200k, 5M, 1.5m, 200000). When the cap is hit:

burnstop: budget exhausted: $1.04 / $1.00 this session (incl. subagents). Halting.

/budget-reset and /budget-disarm work even after a halt — the fuse exempts burnstop's own commands, so it can't block you from clearing it. (You can also cap a whole session at launch: BURNSTOP_BUDGET=$1 claude.)

A Stop hook recomputes spend from Claude Code's own local transcripts each turn (no ledger, nothing written during a run), prices each turn by its model for dollar budgets, and returns continue:false to halt — which overrides /goal so a runaway can't resurrect. It also trips on a structural loop (the same tool call repeated). Full design, pricing table, loop detection, subagent handling, and limitations: docs/architecture.md. Why a hook and not a proxy: docs/spikes.md.

• Python 3.8+, standard library only — no pip install, no build step. (Anyone running Claude Code already has Python.)

Bounded overshoot, not surgical: the fuse is ~one turn behind, and a subagent already mid-run can't be killed (only the parent halts and new spawns are blocked) — set the cap with headroom and bound subagents with maxTurns. Loop detection is a heuristic; the budget cap is the deterministic backstop. On a subscription there's no per-token dollar bill, so a dollar budget is computed from list prices as a proxy for usage burn.

• claude-usage — a local dashboard for Claude Code token usage and cost. burnstop caps your spend; claude-usage shows it.
• pm-skills — the AI operating system for better product decisions (Claude Code plugins).

PRs welcome — see CONTRIBUTING.md. Agent/contributor guide: AGENTS.md.

MIT — see LICENSE.