DEV Community

Javier Fajardo
Javier Fajardo

Posted on

# The Missing Layer of the AI Agent Stack: A Machine-to-Machine Search Engine

#ai #llm #agents #openai

AI agents still search for tools like humans do — parsing READMEs, reading docs, guessing install commands. We built the layer that was missing from every agent stack diagram.

The problem

An AI coding agent needs to send an email. It knows sendgrid exists. What it doesn't know:

  • Is it pip install sendgrid or pip install sendgrid-python?
  • What's the correct import?
  • What parameters does Mail() take?

So it reads the README. Or the docs page. Or hallucinates the API.

A typical library README is 4,000–15,000 tokens. Multiply that by 5 tools in a workflow and the agent has burned half its context window before writing a single line of code.

The question an agent actually needs answered is about 40 tokens:

install: pip install sendgrid
import:  from sendgrid import SendGridAPIClient
task:    send transactional email
Enter fullscreen mode Exit fullscreen mode

What AIAAM does

Two catalogs, one search endpoint:

MAI-1 — executable contracts for installable libraries (PyPI, npm, GitHub)

MAI-API — compressed manifests for remote web APIs (~850 tokens vs thousands for raw OpenAPI)

GET https://aiaam.xyz/api/v1/search?q=send+transactional+email
Enter fullscreen mode Exit fullscreen mode

Response (truncated):

{
  "query": "send transactional email",
  "total": 3,
  "tools": 1,
  "apis": 2,
  "results": [
    {
      "type": "tool",
      "identity": { "aid": "sendgrid-v1" },
      "action": {
        "install_cmd": "pip install sendgrid",
        "execute_cmd": "from sendgrid import SendGridAPIClient"
      },
      "trust": { "reliability_score": 0.91 }
    },
    {
      "type": "api",
      "service_name": "sendgrid_email_activity_api",
      "base_url": "https://api.sendgrid.com",
      "auth_type": "bearer",
      "manifest_url": "https://aiaam.xyz/api/v1/services/sendgrid_email_activity_api/mai-api.json"
    }
  ]
}
Enter fullscreen mode Exit fullscreen mode

Three more examples that work today:

GET /api/v1/search?q=vector+database
# → chroma-v1, weaviate-v1, milvus-v1, faiss-cpu-v1, qdrant-client-v1

GET /api/v1/search?q=audio+transcription
# → openai-whisper-large-v3-v1, whisperx-v1

GET /api/v1/search?q=llm+orchestration
# → langchain-v1, langgraph-v1, crewai-v1
Enter fullscreen mode Exit fullscreen mode

Search is intent-based. "audio transcription" returns whisper because its tags include "audio transcription speech to text stt voice recognition" — not because the aid matches.

The M2M loop 

agent needs tool
  → GET /api/v1/search?q=<intent>
  → receives MAI-1 contract: install_cmd, execute_cmd, input_schema, output_schema
  → executes
Enter fullscreen mode Exit fullscreen mode

No README parsed. No hallucination. One HTTP call.

The same loop works for web APIs via the MAI-API manifest:

agent needs to call Stripe
  → GET /api/v1/services/stripe_api/mai-api.json
  → ~850 tokens: base_url, auth, 20 key endpoints
  → calls Stripe
Enter fullscreen mode Exit fullscreen mode

The raw Stripe OpenAPI spec is ~1.2M tokens. The manifest is ~850. That's a 1,400× reduction.

Where MAI-1 contracts live

We've been submitting pull requests to popular repos to add an AGENTS.md with the contract inline. The key design decision: the contract lives inside the repo, no external dependency.

{
  "aid": "scrapy-v1",
  "logic": {
    "input_schema":  { "type": "object", "properties": { "url": { "type": "string" } } },
    "output_schema": { "type": "array" }
  },
  "trust": {
    "reliability_score": 0.88,
    "latency_ms": 120
  },
  "action": {
    "install_cmd": "pip install scrapy",
    "execute_cmd": "import scrapy"
  }
}
Enter fullscreen mode Exit fullscreen mode

The maintainer owns this block and can edit it. No account, no API key, no external service pinged. The format spreads; the tool doesn't create a dependency.

The earlier approach pointed to aiaam.xyz — maintainers closed those PRs. The inline approach removes the objection entirely.

What's available

  • 135 MAI-1 tool contracts (PyPI, npm, GitHub, Hugging Face)
  • 43 MAI-API manifests (Stripe, Slack, GitHub API, Twilio, Notion, SendGrid, OpenAI, and more)
  • MCP server for Claude Desktop and Cursor: https://aiaam.xyz/mcp/sse 
{
  "mcpServers": {
    "aiaam": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://aiaam.xyz/mcp/sse"]
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Once connected, search_tools("vector database") works as a tool call from any MCP-compatible agent.

Honest notes

  • The catalog covers common tools. If yours isn't listed, you can compile it: POST /api/v1/submit-api with any public OpenAPI URL.
  • MAI-API manifests take the 20 most prominent endpoints from each spec. Complex APIs like Jira have hundreds — the manifest covers the common operations.
  • Contracts are compiled from public sources (PyPI, GitHub, APIs.guru CC0). No affiliation with any library or API provider.
  • Reliability scores are real where we have usage data; placeholders where we don't.

Search endpoint: aiaam.xyz/api/v1/search

API manifests: aiaam.xyz/mai-api

MCP server: https://aiaam.xyz/mcp/sse

Top comments (0)