Cost guard
Security skills that make AI coding agents safe to run. 5 drop-in skills: skill verification, cost tracking, tool authorization, behavioral anomaly detection, and pre-execution safety.
npx -y skills add ArielSmoliar/safe-agent --skill cost-guardAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Per-session cost and token budget tracking for AI coding agents. Monitors cumulative spend, warns at thresholds, and enforces limits with three modes: reject (stop work), alert (warn and continue), downgrade (suggest cheaper approach). Ported from the LOCO-Agent BudgetManager. Trigger phrases: "set a budget", "track my spend", "cost limit", "how much have I spent", "token budget", "spending cap".
SKILL.md
3.8 KB, as published. Nobody here has run it
Cost Guard
Track and limit session costs to prevent runaway spend.
When to Use
- User sets a budget: "/cost-guard $5" or "/cost-guard 100k tokens"
- User asks about spend: "how much have I spent?"
- Proactively, when a session involves many LLM calls, large context windows, or iterative loops that could accumulate significant cost
Budget Configuration
When invoked with arguments, parse the budget:
- Dollar amounts:
/cost-guard $5,/cost-guard $0.50 - Token amounts:
/cost-guard 100k tokens,/cost-guard 500000 tokens - With mode:
/cost-guard $10 alert,/cost-guard 200k tokens reject
Default mode is alert (warn but don't stop).
Three Enforcement Modes
Adapted from LOCO-Agent's BudgetManager:
Reject
Stop work when the budget is exceeded. Report what was spent and what triggered the limit. Ask the user if they want to increase the budget or end the session.
Alert
Warn the user when the budget is exceeded but continue working. Show a clear warning with current spend vs limit. Repeat the warning at each subsequent 25% increment above the limit.
Downgrade
When approaching the budget (>80% consumed), proactively suggest cheaper approaches:
- Shorter responses (skip verbose explanations)
- Fewer tool calls (batch operations instead of one-at-a-time)
- Skip optional validation steps
- Read smaller file sections instead of full files
Cost Estimation
Track cost using these approximations:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Claude Opus | $15.00 | $75.00 |
| Claude Sonnet | $3.00 | $15.00 |
| Claude Haiku | $0.25 | $1.25 |
For each tool call and response in the session, estimate:
- Input tokens: prompt + context + tool results (rough: 1 token ≈ 4 chars)
- Output tokens: assistant response length
- Cumulative total: running sum across the session
These are estimates. State this clearly — actual billing may differ.
Status Report Format
When the user asks about spend, or when a threshold is crossed:
## Cost Guard Status
Budget: $5.00 (alert mode)
Estimated spend: $3.42 (68% of budget)
├─ Input tokens: ~284,000 ($0.85)
├─ Output tokens: ~51,200 ($2.56)
└─ Tool calls: 34
Status: ✓ Within budget
When a threshold is crossed:
⚠ Cost Guard: 82% of $5.00 budget consumed (~$4.10)
Downgrade suggestions:
- Consider reading specific line ranges instead of full files
- Batch remaining edits into fewer tool calls
- Skip re-reading files you've already seen
Continue at current pace? Or adjust budget with /cost-guard $10
Tracking Rules
- Start tracking from the moment the budget is set, not session start
- If no budget is set, do not track or report costs
- Report at these thresholds: 50%, 80%, 100%, and every 25% after
- Include the threshold report inline — don't interrupt the user's task flow, just append a brief status line after your response
- When the session is clearly ending (user says thanks, goodbye, etc.), include a final cost summary
Limitations
Be transparent: this is estimation, not metering. State clearly that:
- Token counts are approximated from character length
- Actual costs depend on prompt caching, batching, and pricing changes
- This is a planning tool, not a billing tool