1
# Tool Budget Guard
2
3
An agent with no budget does not fail loudly, it fails expensively. It retries the
4
same broken command nine times, re-reads a file it already has, and burns twenty
5
minutes before anyone notices. This harness puts a ceiling on tool calls, tokens and
6
wall-clock time, and, the part most budget code skips, says exactly what happens when
7
the ceiling is hit.
8
9
## When to use
10
11
- Any long-running or autonomous task where nobody is watching each step.
12
- Agents that call paid APIs, where a loop is a bill.
13
- Any agent that has ever gotten stuck repeating itself.
14
15
## Install
16
17
```bash
18
npx openagents-cli add openagents/tool-budget-guard
19
```
20
21
| Runtime | Installed to |
22
|---|---|
23
| `claude-code` | `.claude/skills/tool-budget-guard/` |
24
| `codex` | `.codex/skills/tool-budget-guard/` |
25
| `openai-agents` | `agents/tool-budget-guard/` |
26
| `langgraph` | `graphs/tool-budget-guard/` |
27
| `generic` | `.openagents/tool-budget-guard/` |
28
29
## What is in the package
30
31
- `HARNESS.md` - the budget model, the checks, and the escalation ladder.
32
- `budgets.md` - starting numbers by task shape, and how to tune them from real runs.
33
34
## License
35
36
MIT
37