The Sugra API Skills

You are writing a module. It needs three live things at once: FX rates, weather at the destination, and which ports are in play. You tell the agent to wire Sugra API. Without a skill it will guess a path, skip the envelope, and paste a number it saw on the web. With The Sugra API Skills it already knows the map: Sugra Finance for FX, Sugra Earth for weather and maritime ports, one key, the same envelope on every call, and the source plus the observation time in the answer. It confirms the operation on docs.sugra.ai instead of inventing a route. The figure it returns is current because it came from the API, not from a page that ranked this morning.

That is the why. The rest is what the pack actually is.

Your agent already knows how to call a URL and how to use a tool. It does not know Sugra API. Left alone that is not a model failure. It is missing procedure.

A skill is that procedure, written down. A folder with a SKILL.md: when to load it, how to authenticate, how to find the operation, how to call it, how to read the envelope and name the source. The agent loads the description at startup and the rest only when the task matches, so the context window pays for what it uses. Anthropic shipped the format as Agent Skills; it is now an open standard. MCP hands the agent tools. A skill encodes when and how to use them, or how to go straight over HTTPS if the host has no MCP at all.

The official pack is seven skills, one plugin sugra-api, MIT. It is not the MCP server (sugra-api-mcp) and not the HTTP cookbook (sugra-api-cookbook). Those are the plug and the recipes. The skill pack is the teacher.

Sugra API is one HTTP API, one key, seven directions (Sugra Finance, Sugra Macro, Sugra Entity, Sugra Net Atlas, Sugra News, Sugra Earth, Sugra Research), 1,500+ endpoints, 160+ primary sources, 36 domains. The pack does not replace the docs. It tells the agent to use them.

Published in the OpenAI Skills Directory

The Sugra API Skills pack is published in the official OpenAI Skills Directory for ChatGPT and Codex. That is the listing to install, not a side note.

Two complete ways in

HTTPS to https://sugra.ai with x-api-key. MCP on the hosted connector mcp.sugra.ai, or locally with pip install sugra-api-mcp. Both reach the same API. The skills treat them as complete paths. Use the surface the host already has, or the one you named.

The same pack covers two kinds of work.

Writing software. The agent is in your repo, adding the module above, a client, a job. discover-and-call forbids inventing routes: confirm the operation, then GET or POST with the key, or search_endpoints / describe_endpoint / call_endpoint on MCP. envelope-and-attribution is the contract every direction shares: {data, meta}, cite the source and data_time. auth-and-quota is 401 and 429, Bearer versus x-api-key, and not logging the key. The code it writes still has to be correct after the session ends.

Asking a question. You want the same three numbers in chat, not in a module. The same skills fire. The agent does not answer from a snippet. It finds the operations, calls them, and cites. cross-domain-briefing keeps it to two or three directions in one answer, not a mash of seven. live-docs is where to look when a catalog on disk might lag the live OpenAPI.

We wrote the architecture version of this last winter: skills are procedural memory, checked into git. This pack is that file for Sugra API.

What the pack actually contains

Skill	What it teaches
`using-sugra-api`	Map of surfaces and the seven directions. Which skill to load next.
`live-docs`	docs.sugra.ai, OpenAPI, `/sources`, `/stats`, the blog.
`connect`	HTTPS, hosted MCP, stdio, self-host, each client.
`auth-and-quota`	Key, plans, 401/429, Bearer versus `x-api-key`.
`discover-and-call`	Find and call an operation on HTTP or MCP. Do not invent the path.
`envelope-and-attribution`	`{data, meta}`, source names, clocks.
`cross-domain-briefing`	Two or three directions in one answer.

Author: Sugra Systems, Inc. Format: Agent Skills. Get a key at app.sugra.ai/settings/billing (Free: 50 requests/day).

Install

Claude Code:

bash

claude plugin marketplace add \
  Sugra-Systems/sugra-api-skills
claude plugin install \
  sugra-api@sugra-api-skills

Grok:

bash

grok plugin marketplace add \
  Sugra-Systems/sugra-api-skills
grok plugin install sugra-api --trust

Codex copies plugins/sugra-api/skills/ into ~/.agents/skills/. Cursor and Gemini CLI do the same into their own skills directories. ChatGPT and Codex: the official OpenAI Skills Directory. Hosted MCP is mcp.sugra.ai. Local MCP is pip install sugra-api-mcp.

The README in the repo is the install page. This post is why you would.

A skill does not make the model smarter. It stops the model from improvising the one thing it should not: where FX, weather and ports actually live, and how to call them today.