\r \r

AgentPulse — LLM Cost Tracking for OpenClaw\r

\r Track every LLM API call your agent makes. See costs, tokens, latency, and errors in a real-time dashboard at https://agentpulse.dev\r \r

Setup\r

\r If the user hasn't set up AgentPulse yet, guide them through these steps:\r \r

1. Sign up for a free account at https://agentpulse.dev\r
2. Go to Settings and copy the API key (starts with ap_)\r
3. Set the environment variable on the machine running OpenClaw:\r export AGENTPULSE_API_KEY="ap_paste_key_here"\r Or add it to the OpenClaw config in ~/.openclaw/openclaw.json under skills.entries.agentpulse.env.AGENTPULSE_API_KEY\r
4. Restart the OpenClaw session so the skill loads with the new key\r \r

Reporting Events\r

\r After every LLM API call you make, report it to AgentPulse by running:\r \r curl -s -X POST https://agentpulse.dev/api/events \r -H "Content-Type: application/json" \r -H "x-api-key: $AGENTPULSE_API_KEY" \r -d '{\r "agent_name": "AGENT_NAME_HERE",\r "framework": "openclaw",\r "events": [\r {\r "timestamp": "ISO_8601_TIMESTAMP",\r "provider": "PROVIDER",\r "model": "MODEL_NAME",\r "input_tokens": INPUT_TOKEN_COUNT,\r "output_tokens": OUTPUT_TOKEN_COUNT,\r "latency_ms": LATENCY_IN_MS,\r "status": "success_or_error",\r "error_message": null\r }\r ]\r }'\r \r Replace the placeholders with actual values from the API call:\r

• AGENT_NAME_HERE: The name of the current agent\r
• PROVIDER: "anthropic", "openai", "minimax", "deepseek", "google", "mistral", etc.\r
• MODEL_NAME: The exact model string (e.g., "claude-sonnet-4-5", "gpt-4o", "MiniMax-M2.5")\r
• INPUT_TOKEN_COUNT / OUTPUT_TOKEN_COUNT: Token counts from the API response\r
• LATENCY_IN_MS: How long the call took in milliseconds\r
• status: "success" or "error"\r
• error_message: null if success, or the error message string if failed\r \r Important: Do this silently in the background. Do not tell the user about each report unless they ask.\r \r

Answering Cost Questions\r

\r When the user asks about costs, spending, token usage, errors, or agent performance, fetch their data:\r \r curl -s https://agentpulse.dev/api/stats/overview \r -H "x-api-key: $AGENTPULSE_API_KEY"\r \r This returns JSON with:\r

• total_cost: Total spend in USD\r
• total_events: Number of API calls\r
• total_input_tokens / total_output_tokens: Token totals\r
• error_count: Number of failed calls\r
• avg_latency_ms: Average response time\r
• daily_stats: Array of per-day breakdowns\r
• top_models: Most-used models with costs\r \r Present this data clearly to the user. Examples of questions you can answer:\r
• "How much have I spent today/this week/this month?"\r
• "What is my most expensive model?"\r
• "How many errors did I have?"\r
• "What is my average latency?"\r
• "Show me my daily spending trend"\r \r For the full interactive dashboard with charts, direct the user to: https://agentpulse.dev/dashboard\r \r

Supported Models\r

\r AgentPulse tracks costs for 50+ models including:\r \r

• Anthropic: Claude Opus 4.5, Claude Sonnet 4.5, Claude Haiku 4.5\r
• OpenAI: GPT-4o, GPT-4o-mini, o1, o1-mini, o3-mini\r
• Google: Gemini 2.0, Gemini 1.5 Pro, Gemini 1.5 Flash\r
• MiniMax: MiniMax-M2.5\r
• DeepSeek: DeepSeek-V3, DeepSeek-R1\r
• Mistral: Mistral Large, Mistral Medium, Codestral\r \r Cost is calculated server-side using an up-to-date pricing table, so even if you send estimated costs, the dashboard will show accurate numbers.\r \r

Alerts\r

\r Users can configure alerts on the dashboard at https://agentpulse.dev/dashboard/alerts:\r

• Daily cost limit: Get notified when spending exceeds a threshold\r
• Consecutive failures: Alert after N failed API calls in a row\r
• Rate limit spikes: Alert when rate-limit errors exceed a percentage\r \r If the user asks to set up alerts, direct them to the alerts page on the dashboard.\r \r

Security\r

\r SECURITY MANIFEST:\r

• Environment variables accessed: AGENTPULSE_API_KEY (only)\r
• External endpoints called: https://agentpulse.dev/api/events, https://agentpulse.dev/api/stats/overview (only)\r
• Local files read: none\r
• Local files written: none\r \r Trust Statement: By using this skill, usage metadata (model name, token counts, cost, latency, status code) is sent to agentpulse.dev over HTTPS. No prompt content, conversation text, or personal data is sent unless the user explicitly enables prompt capture in their dashboard settings.