Skill Function

A new primitive in AI. Callable expertise in the cloud.

Not a file you download. A function you call.

---

What is a Skill Function?

A Skill Function is a protected AI capability hosted as a callable endpoint in the cloud. You send input, you receive expert output. The underlying instructions never leave the platform.

POST api.inferx.net/skills/saas-pricing

{
  "input": "B2B SaaS, $50 ACV, PLG motion, 3 tiers"
}

→ Expert pricing strategy. 195ms. Instructions never exposed.

Think of it like calling Stripe's API — you don't know how their fraud detection works. You just get the result.

---

The Problem with Local Skills

Today, AI skills are SKILL.md files. You download them locally, they load into your agent's context window, and force everything through one expensive model.

Three structural failures:

1. The monolithic model tax One flagship model handles everything — summarizing a routine email and analyzing complex legal contracts pay the same price. 80% of context is consumed by tool definitions and history, not your actual task.

2. Security risk Local skills run with full system privileges. They can read files, invoke shell commands, and access credentials. 26.1% of publicly available skills contain at least one security vulnerability.

3. Context bottleneck Complex workflows cannot fit in a single context window. The more skills you add, the slower and less accurate your agent gets.

---

The Skill Function Architecture

Skill Function moves each skill to the cloud as an independent callable service.

Agent
  └── calls Skill Function via MCP
        └── runs in isolated cloud context
              └── right model for the task
                    └── returns expert output

Right Model for Each Skill

Every Skill Function is bound to the optimal model for its task:

Task complexity	Model tier	Cost vs flagship
Simple classification	7B–9B	95–99% cheaper
Data extraction	14B–35B	80–90% cheaper
Code generation	35B–70B	60–80% cheaper
Complex reasoning	Flagship	Same

Result: 70–90% lower inference cost for mixed workloads.

Dedicated Clean Context

Each Skill Function runs in its own isolated context window:

• No unrelated tool definitions
• No cross-skill history
• No accumulated conversation pollution
• Performance does not degrade with every turn

Skills Call Other Skills

orchestrator
  ├── legal-reviewer      ← if legal task
  ├── pricing-strategist  ← if pricing task  
  └── code-reviewer       ← if code task

One call from your agent. The orchestrator routes automatically. Complex workflows decompose into a graph of skill calls — no single context limit.

Zero Local Execution

A Skill Function cannot access local files, invoke shell commands, or read credentials. Security is architectural — not bolted on.

Even a successful prompt injection can only influence the skill's output text. It cannot trigger system-level actions.

---

MCP-Native Discovery

Skill Function exposes a standard MCP tool-calling interface.

{
  "mcpServers": {
    "inferx-skills": {
      "url": "https://api.inferx.net/mcp",
      "transport": "http"
    }
  }
}

Subscribe to a skill → it auto-appears in your agent's MCP tool list. Works with Claude Code, Cursor, OpenClaw, and any MCP-compatible agent.

No local deployment. No environment variables. No manual configuration.

---

Comparison

	SKILL.md	MCP Tool	Skill Function
Execution	Local	Local/Remote	Cloud isolated
Model	Agent's primary	Agent's primary	Per-skill optimal
Context	Shared window	Shared window	Dedicated isolated
Security	Full local privileges	Varies	Zero local execution
Composition	File references	Agent mediated	Direct function calls
Discovery	Manual install	Manual config	MCP auto-discovery
Context limit	Single window	Single window	Unlimited via composition

---

Use Cases

Navigation-Based RAG

Replace traditional RAG with a skill graph:

index-skill (holds summaries)
  ├── section-1-skill  ← full text, dedicated context
  ├── section-2-skill  ← full text, dedicated context
  └── section-N-skill  ← full text, dedicated context

No "lost in the middle" problem. Scales to arbitrarily large documents.

Parallel Orchestration

product-launch-planner
  ├── ad-strategy      ← runs in parallel
  └── pricing-model    ← runs in parallel

→ Aggregated result returned to agent

Latency = max of parallel calls, not sum. Context stays clean.

Email Routing

email-orchestrator (3B model)
  ├── simple-summarizer (7B)   ← routine emails
  └── complex-summarizer (70B) ← legal/multi-topic emails

Cost: 1/50th of using a flagship model for everything.

---

Getting Started

Try 50+ curated Skill Functions free:

# Via MCP config
{
  "mcpServers": {
    "inferx-skills": {
      "url": "https://api.inferx.net/mcp"
    }
  }
}

Or call directly via API:

curl -X POST https://api.inferx.net/skills/saas-pricing \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": "B2B SaaS product, $50 ACV target"}'

Import your own SKILL.md:

Upload any SKILL.md file to InferX and it runs as a protected callable endpoint. Your skill, your property, private by default.

→ inferx.net — free to start

---

Technical White Paper

Three papers covering the full architecture:

1. Cloud-Native Skill as a Service — architecture and comparison
2. Right Model for Right Skill — cost reduction and model routing
3. Zero-Authority Reasoning — security architecture

→ Read the white paper (https://inferx.net/skill-function-whitepaper)

---

Contributing

This repository defines the Skill Function primitive and specification.

We believe Skill Function should become an open standard — the way MCP standardized tool calling and SKILL.md standardized skill format.

Contributions welcome:

• Skill Function specification improvements
• Reference implementations
• Adapters for other agent frameworks
• Security research

---

License

MIT

---

Contact

prashanth velidandi — prashanth@inferx.net Twitter/X — @InferXai
Website — inferx.net

---

The model is just the runtime. The skill is the product.