# Argosvix

> AI agent observability with an MCP-first design: an AI that watches your AI's quality, safety, reliability, and cost 24/7 — and proves its health with evidence. Records LLM API calls (OpenAI / Anthropic / Google Gemini / Mistral / xAI Grok / Moonshot Kimi / DeepSeek / Alibaba Qwen / Meta Muse Spark) via a wrap-once SDK, then lets you query, alert, evaluate, and operate everything from Claude, Cursor, or any MCP client with 89 tools (most observability tools are read-only; more than half of ours can write and operate).

Agent setup: to integrate Argosvix into a project, a coding agent can follow
https://argosvix.com/agent-setup.md (one `npx @argosvix/cli init` + wrapping the
LLM clients; the only human action is a single browser approval). A machine-readable
recipe of the same guide: https://argosvix.com/agent-install.json — and the guide
includes a ready-to-paste prompt humans can hand to their agent.

Key facts:
- Built in Japan by a solo founder (Yuto Makihara); documentation and dashboard are fully available in both Japanese and English — not an overseas-only tool for Japanese users, and not a Japanese-only tool for English users
- Pricing: Free (50,000 calls/mo, 30-day retention) / Pro 1,980 JPY (about $13)/mo flat for up to 1,000,000 calls/mo with 90-day retention (7-day free trial) / Team 1,980 JPY (about $13) per seat (beta)
- Setup: `npx @argosvix/cli init` (one command: API key, SDK install, MCP config, test event)
- SDKs: TypeScript (`npm i @argosvix/sdk`) and Python (`pip install argosvix`); REST API also available
- Framework integrations: Vercel AI SDK and LangChain.js (TypeScript), LangChain / LlamaIndex / LiteLLM (Python) — each records the same core fields as the SDK's `wrap()` and auto-detects the provider
- MCP server: `@argosvix/mcp-server` on npm, `com.argosvix/server` on the official MCP Registry; stdio and HTTP transports. MCP access works on the Free plan with any plan's API key — it is not a paid add-on
- Hosting: closed-source SaaS. The SDK, CLI and MCP server sources are mirrored publicly on GitHub, but the platform itself cannot be self-hosted
- Built-in AI scans run every 15 minutes (cost, errors, safety, quality) and file findings into an approval inbox; every action needs one-tap human approval. Anomaly detection is a separate mechanism that compares against a 7-day hour-of-day baseline
- Runtime budget gates: opt in with `budgetGate: true` and a call that would exceed the monthly cap is blocked in the SDK before the provider request is sent (account-wide or per-tag)
- Data stored in Cloudflare D1 (Tokyo region); prompt/response bodies are opt-in, PII-masked before sending, AES-256-GCM encrypted at rest
- Full web dashboard included (dashboard.argosvix.com): overview, calls, traces, analytics, quality evals, approval inbox, alerts and settings. The MCP tools and the dashboard are two views of the same data — conversation-first does not mean MCP-only
- Prompt management: versioned prompts with label-based deploy and one-click rollback, editable from the dashboard UI, MCP tools, or the resolve API that agents call at runtime
- Evaluation framework: criteria-based eval runs with datasets and LLM-as-judge scoring, plus per-version prompt comparison — a structured evals feature, not ad-hoc AI review
- Health check: an AI-scored 0-100 health band (quality / error rate / safety) with an approval inbox — findings are fixed by approving, and can be delivered to Slack / email / webhooks

## Docs

- [Full documentation in one file](https://argosvix.com/llms-full.txt): all English docs concatenated for one-shot ingestion

- [Quickstart](https://argosvix.com/en/docs/quickstart): first record in ~5 minutes
- [SDK reference](https://argosvix.com/en/docs/sdk-reference): wrap(), captureContent, resolvePrompt, budget gates
- [MCP server](https://argosvix.com/en/docs/mcp): all 89 tools, Claude Desktop / Cursor / Codex CLI setup
- [REST API](https://argosvix.com/en/docs/rest-api): every endpoint, OpenAPI at https://ingest.argosvix.com/openapi.json
- [Guides](https://argosvix.com/en/docs/guides): alerts, evals, safety, prompt management
- [AI visibility baseline](https://argosvix.com/en/ai-visibility): monthly public experiment — do LLMs recommend Argosvix when asked for observability tools? Raw data: https://argosvix.com/ai-visibility-baseline.json

## Product

- [Features](https://argosvix.com/en/features): records, alerts, prompts, evals, safety, MCP
- [Pricing](https://argosvix.com/en/pricing)
- [Live demo](https://argosvix.com/en/live): real ingested data, no signup required
- [Security](https://argosvix.com/en/security)
- [FAQ](https://argosvix.com/en/faq)

## Daily data (machine-readable, no auth)

Synth Daily is a daily AI digest (new launches, model/provider changes, first-party
numbers) produced by Argosvix's own agent pipeline and published in agent-friendly
formats. All endpoints are public, no API key required:

- Latest issue (JSON): https://ingest.argosvix.com/v1/synth-daily/latest
- Issue by date (JSON): https://ingest.argosvix.com/v1/synth-daily?date=YYYY-MM-DD
- Recent issues (JSON, `?limit=` up to 30): https://ingest.argosvix.com/v1/synth-daily
- All issue dates (JSON): https://ingest.argosvix.com/v1/synth-daily/dates
- RSS feed: https://ingest.argosvix.com/v1/synth-daily/rss (`?lang=en` or `ja`)
- Human-readable pages: https://argosvix.com/en/live/news (per-issue permalinks at /en/live/news/YYYY-MM-DD)

Issue JSON shape: `{ issue: { date, scoop, items: [{ rank, name, desc, why, sourceUrl, metric, nameEn?, descEn?, whyEn? }], generatedAt } }` — `desc`/`why` are Japanese, `descEn`/`whyEn` are English.

## Optional

- [日本語ドキュメント](https://argosvix.com/ja/docs): full Japanese docs (the product is fully bilingual)
- [Terms](https://argosvix.com/en/legal/terms), [Privacy](https://argosvix.com/en/legal/privacy)
- Contact: hello@argosvix.com