Shingou skills

Plug the Shingou point-in-time crypto market sentiment API into Claude or any AI agent. This repo is a skill that teaches the agent to use the API correctly (auth, symbols, quotas, and honest framing of an hourly signal), plus MCP tools for direct access.

Works instantly with a free API key (shingou.io/dashboard, no card). The paid key is what removes the 24h delay outside the live majors, and what buys history depth.

Path	What it is
plugins/shingou/skills/shingou-api/	The skill: SKILL.md + full API reference + worked flows
MCP endpoint	https://api.shingou.io/mcp — the same API as callable tools
src/stdio-proxy.mjs	Packaged stdio server, for hosts that launch a local process instead of connecting to a remote one

Install

Claude Code (plugin)

/plugin marketplace add shingou-io/shingou-skills
/plugin install shingou@shingou-skills

Then export your key so the agent can use it:

export SHINGOU_API_KEY=sk_live_...

claude.ai (skill upload)

Zip the skill folder (folder at the zip root) and upload it under Settings → Capabilities → Skills:

cd plugins/shingou/skills && zip -r shingou-api.zip shingou-api

There are no environment variables on claude.ai — the skill will ask you for the API key in-conversation.

Any other agent

Point your agent at plugins/shingou/skills/shingou-api/SKILL.md (plain markdown, no Claude-specific machinery), or at the live LLM-oriented reference: shingou.io/llms-full.txt.

MCP

Connect the Shingou API as tools (get_sentiment, get_sentiment_history, get_events, get_symbols) over Streamable HTTP. The Claude Code plugin configures this automatically — with SHINGOU_API_KEY exported, the shingou MCP server is ready after /plugin install. To connect without the plugin:

claude mcp add --transport http shingou https://api.shingou.io/mcp \
  --header "x-api-key: sk_live_..."

Requests made through MCP are attributed server-side — no User-Agent handling needed. claude.ai remote connectors (OAuth) are planned.

Packaged server (stdio)

Streamable HTTP is the way in, and every client above uses it. Two places cannot: hosts that only launch a local process, and directories that index a server by building it and speaking MCP to its stdin. src/stdio-proxy.mjs bridges both to the same remote endpoint. No dependencies, one HTTP POST per JSON-RPC message.

Published to npm, so a host that launches a local process needs no clone:

claude mcp add shingou -- npx -y @shingou-io/mcp-stdio

No build step, no dist/, no dependencies: the entrypoint is plain JavaScript and runs on Node 18 or newer. It was TypeScript until 0.3.1, which cannot work from an installed package, because Node refuses to strip types for anything under node_modules on every version.

From a clone or an image instead:

SHINGOU_API_KEY=sk_live_... node src/stdio-proxy.mjs

docker build -t shingou-mcp-stdio .
docker run -i --rm -e SHINGOU_API_KEY=sk_live_... shingou-mcp-stdio

SHINGOU_MCP_URL overrides the endpoint. The key is optional: initialize, tools/list and get_symbols all answer without one, so a client can discover the tools before the user has a key. The data tools need it.

Two layers of checks, split by what a failure would mean:

node --test test/proxy.test.ts   # hermetic: stub upstream, no network, no key
scripts/smoke.sh                 # live: builds the image, talks to the real endpoint

The first is what CI runs, verbatim and with no package manager, because it fails only when the bridge regresses. scripts/smoke.sh runs on a daily schedule instead — it can go red because the API moved or a deploy is mid-flight, and that should not block a pull request.

Quota math (free tier)

1,000 requests/day, 30/min burst. The signal changes once per hour per asset, and one /v1/sentiment call batches up to 50 symbols — so even polling the full universe every hour costs 24 requests a day. There is no reason to poll faster.

Free history reaches back one day, which is enough for /v1/history/sentiment to return a real series and not enough to backtest on. Depth is the paid axis: 90 days on starter, 365 on quant, 730 on pro.

What the signal is (and is not)

Point-in-time first. Every history bucket carries an as-of timestamp and a reconstructed flag, and the bucket's hash is committed to a public append-only log at publish time. So no lookahead is something you can check, not something we ask you to believe. The log is at shingou-io/shingou-signal-log.

The signal itself is one per asset per hour: score ∈ [-1, 1], confidence ∈ [0, 1], a direction call, dominant event types, and the source articles behind it. Measured performance, including the negative results, is published at shingou.io/research. The honest uses are filter, sizing, and kill-switch. Never an entry generator on its own. The skill instructs agents to disclose signal age (free-tier delays) and reconstructed history buckets rather than hide them.

License

MIT — see LICENSE. The skill and MCP tools are free distribution; the subscription is the data. API usage is governed by the Shingou terms (commercial use included on paid plans; free tier is for personal/evaluation use).

Not investment advice. This is an integration layer for a signal, not trading recommendations; backtest before risking anything.