AI agent operation
From reading to alerts, budgets, prompts, evals, and approvals — operate everything by conversation with your AI agent. Works with a Free-plan API key.
Sample conversation in Claude Desktop
Is Claude getting slow lately? How does it compare to the others?
→ get_percentiles(metric=latency, p=95, since=7d, groupBy=model)
gpt-5.5: 1.8s, claude-sonnet-5: 1.2s, gemini-3.5-flash: 0.9s (example). Claude is the fastest.
Alert me on Slack when Claude fails over 8%.
→ create_alert(name=Claude error rate, condition=model=claude-*, error_rate(15m) > 8%, channel=slack)
Alert "Claude error rate" created. Slack will be notified if a 15-minute window exceeds 8%.
89 MCP tools
From reading to creating, changing, and approving. Two anonymous tools (get_synth_daily / get_daily_readthrough) work without a key.
Read39query_calls · get_cost_summary
Create12create_alert · create_prompt
Update9update_prompt · rename_project
Delete12delete_alert · bulk_delete_calls
Operate17run_eval · acknowledge_alert
Anonymous tools, no key needed get_synth_daily get_daily_readthroughno API key
Works with major agents
Claude Desktop / Cursor / Codex CLI / Anthropic API / OpenAI tools. Available from the official MCP registry and the Claude, Cursor, and Gemini marketplaces.
Claude Desktop
Cursor
Codex CLI
Anthropic API / OpenAI Tool
Audited operations, fewer tokens
Write operations land in the audit log with the actor recorded. Replacing UI round-trips with tool calls also keeps investigation token usage down.
Example
audit.log
09:41:12update_promptactor: mcp · argk_…7f2
09:29:05acknowledge_alertactor: dashboard
09:14:47create_alertactor: mcp · argk_…7f2
✓ Every write operation is logged