argosvix/mcp-server

Argosvix MCP server lets AI agents (Claude Desktop, Cursor, Codex CLI, custom MCP clients) query, manage, and operate their LLM observability data directly from the conversation. Supports both **stdio** (subprocess) and **HTTP** (remote / self-host) transports.
**Surface:** 89 tools (86 generally available + 3 internal operations tools that return 403 for customer accounts) / 3 resources / 8 resource templates / 3 prompts. Health and anomaly endpoints (get_account_health / detect_anomaly / propose_alert_rules / classify_calls_batch / propose_eval_criteria) plus a runtime control plane (budget gates / policy gates / human-approval gates) let an agent both observe and act. Release history is available on npm.
Why
You're already sending LLM calls through @argosvix/sdk or the Python SDK. Now ask Claude / Cursor questions like:
- "What was my OpenAI cost over the past 24h?"
- "Which alerts are firing right now?"
- "Show me the 20 most expensive calls today."
No dashboard tab-switching. The agent fetches the data via this MCP server using your Argosvix API key.
Install
**One-click** (starts keyless; set ARGOSVIX_API_KEY afterwards to use the tools for your own account):
Claude Code users can install the argosvix plugin instead — it bundles this server plus setup skills:
**Manual:**
Configure (Claude Desktop)
Edit your Claude Desktop config:
- macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
- Windows: %APPDATA%\Claude\claude_desktop_config.json
Restart Claude Desktop. All 89 tools appear under the argosvix__ prefix (e.g. argosvix__query_calls, argosvix__get_account_health, argosvix__create_budget_gate).
> **No API key yet?** The server also starts without ARGOSVIX_API_KEY in introspection-only mode: all 89 tools are listed so you can evaluate the surface, and two public data tools (get_synth_daily / get_daily_readthrough — Argosvix's own daily AI digest and OSS readthrough) work with no key at all. Every other tool call returns instructions for getting a key at https://dashboard.argosvix.com/api-keys.
>
> Notes on the public data tools: they are part of the default full profile only (not in ARGOSVIX_MCP_PROFILE=core), and like every other tool they honor ARGOSVIX_API_BASE — if you point the server at a different base, they fetch /v1/synth-daily / /v1/readthrough from that base without authentication.
Configure (Cursor)
Edit ~/.cursor/mcp.json:
Tool profile (ARGOSVIX_MCP_PROFILE)
The default profile is full (all 89 tools). If your MCP client's context budget is tight, set ARGOSVIX_MCP_PROFILE=core to expose only the 11 essentials for day-to-day operations:
query_calls · aggregate_calls · get_cost_summary · get_percentiles · get_account_health · detect_anomaly · list_alerts · create_alert · silence_alert · unsilence_alert · get_deployed_prompt
Works in both stdio and HTTP transports. Unknown values fall back to full with a warning on stderr.
Description language (ARGOSVIX_MCP_LANG)
Tool, resource, and prompt descriptions returned by tools/list / resources/list /
prompts/list are in English by default. Set ARGOSVIX_MCP_LANG=ja to get the original
Japanese descriptions instead. Unset or unknown values fall back to en (with a warning
on stderr for unknown values). Tool names, input schemas, and behavior are identical in
both languages — only the human/agent-facing description text changes.
Works in both stdio and HTTP transports.
Tools (highlights)
89 tools in total — the full list is returned by tools/list. A sample of the core read / write surface:
| Tool | Purpose | Type |
|---|---|---|
| query_calls | Recent LLM call records, filterable by provider / model / time range | read |