Agent skills - concentrate

AI coding agents guess at API shapes they don’t know. The Concentrate skills library replaces the guess with documentation-grounded context: install it once, and your agent writes Responses API calls, picks models from the live catalog, and respects data-control boundaries without you pasting docs into the chat. Skills are contextual and load automatically. When a request matches a skill’s triggers, your agent pulls in that skill and applies it; the rest of the time it stays out of the way.

Skills follow the open Agent Skills standard, so the same library works across every agent that implements it. For agents with a plugin system, install through the plugin so skills auto-update.

Quick start

Claude Code
GitHub CLI
Cursor
OpenCode

/plugin marketplace add concentrate-ai/skills-library
/plugin install concentrate@concentrate

Run both commands inside Claude Code. Confirm the install with /plugin list.

Install every skill in the library:

gh skill install concentrate-ai/skills-library

Or install just one:

gh skill install concentrate-ai/skills-library concentrate-models

Go to Settings → Rules → Add Rule → Remote Rule (GitHub) and enter concentrate-ai/skills-library.

mkdir -p ~/.config/opencode/skills
git clone https://github.com/concentrate-ai/skills-library.git /tmp/concentrate-skills
cp -r /tmp/concentrate-skills/skills/* ~/.config/opencode/skills/
rm -rf /tmp/concentrate-skills

Copied skills don’t auto-update. Re-run the commands to pick up changes.

Set your API key in the environment so the agent never writes a literal key into your source:

export CONCENTRATE_API_KEY="sk-cn-v1-..."

What’s in the library

Six skills, each scoped to one area of the platform and grounded in the pages linked below.

Skill	Covers	Docs
`concentrate-responses`	Responses API requests, stateful turns, function tools, structured JSON, web search, SSE streaming, caching, errors	Create response
`concentrate-models`	Live catalog discovery, provider prefixes, capabilities, pricing, auto routing, fallbacks	List models
`concentrate-multimodal`	Image inputs, base64 and URL sources, per-provider format and resolution limits	Multi-modal
`concentrate-data-controls`	Zero Data Retention, request logging, PII redaction with `redact-v1`, BYOK routing and billing	ZDR
`concentrate-integrations`	Client setup for Claude Code, Claude Desktop, Cursor, and OpenAI-compatible SDKs, plus gateway migrations	Integrations overview
`concentrate-alerts`	Spend anomalies, error spikes, entity limits, dormant keys, balance alerts, recurring reports	Alerts overview

What the skills prevent

Each skill carries explicit boundaries, so the agent avoids the mistakes that are easy to make against a multi-provider gateway.

Wrong tool schema. Responses function tools use top-level name and parameters. The nested Chat Completions function shape is rejected with a 400, and the skill keeps the agent on the correct one.
Doubled base paths. Anthropic-native clients append /v1 themselves, so they take https://api.concentrate.ai. Passing the /v1 URL produces /v1/v1/... and 404s.
Stale model lists. The catalog changes. Skills query it live instead of recalling a model table from training data.
Capability assumptions. Image, PDF, tool, structured-output, and ZDR support are properties of a model and provider pair, never inferable from a model family name.
Invented endpoints. Agents are told to use only documented surfaces rather than inventing embeddings, batch, or fine-tuning routes.

Live model queries

concentrate-models ships a dependency-free Python script that reads the public catalog endpoints. It needs no API key, so an agent can explore models before you’ve configured anything.

python3 <skill-path>/scripts/concentrate_models.py list --search claude
python3 <skill-path>/scripts/concentrate_models.py list --capability input.image --sort context
python3 <skill-path>/scripts/concentrate_models.py show gpt-4o
python3 <skill-path>/scripts/concentrate_models.py compare gpt-4o claude-sonnet-4-6
python3 <skill-path>/scripts/concentrate_models.py providers

Add --json when the output feeds another program.

show reports per-provider context windows, pricing, and ZDR status for one model. That per-provider view is what makes ZDR decisions correct, since the combined list models response omits the zdr field.

Example usage

Ask in plain language. The agent loads the matching skill and applies it. “Set up a Concentrate call with a weather tool.” concentrate-responses loads, and the agent produces the correct Responses tool shape:

curl https://api.concentrate.ai/v1/responses \
  -H "Authorization: Bearer $CONCENTRATE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o",
    "max_output_tokens": 200,
    "input": "What is the weather in Paris? Use the tool.",
    "tools": [{
      "type": "function",
      "name": "get_weather",
      "description": "Get weather for a city",
      "parameters": {
        "type": "object",
        "properties": { "city": { "type": "string" } },
        "required": ["city"]
      }
    }]
  }'

“Which models support image input with the largest context?” concentrate-models loads, and the agent queries the live catalog rather than answering from memory:

python3 <skill-path>/scripts/concentrate_models.py list --capability input.image --sort context

“Can I use this model under ZDR?” concentrate-data-controls loads, and the agent inspects the specific model and provider pair, accepting only providers whose zdr field is an object.

Reading the output

Responses can contain reasoning and tool-call items alongside text, so extract by item type rather than by array position:

function outputText(response) {
  return response.output
    .filter((item) => item.type === "message")
    .flatMap((item) => item.content ?? [])
    .filter((part) => part.type === "output_text")
    .map((part) => part.text)
    .join("");
}

The response also reports the resolved model, usage, and cost. Log those when routing or spend matters; the resolved model tells you which provider actually served the request.

Repository

The library is open source at concentrate-ai/skills-library under the MIT license. Skill instructions are grounded in these docs, so corrections here flow into agent behavior. If a skill produces wrong or outdated guidance, open an issue on the repository or email support@concentrate.ai.

​Quick start

​What’s in the library

​What the skills prevent

​Live model queries

​Example usage

​Reading the output

​Repository

Quick start

What’s in the library

What the skills prevent

Live model queries

Example usage

Reading the output

Repository