Looking for first 10 testers

Proof

Code patterns, not hallucinations.

Your AI coding assistant, backed by patterns from real library source code.
"I'll build it. Using what works."

terminal — proof-mcp
$curl -I https://mcp.aigentys.com
HTTP/2 200
server: proof-mcp
$# Add to your MCP config, restart, and ask:
$"List the available MCP tools."
→ 5 tools: catalog, search, explain, build_agent, report

What your AI assistant does today

  • Generates code from training data
  • Sometimes hallucinates APIs that don't exist
  • Has no idea if a pattern actually works in production
  • Trained months ago — the library changed last week

What Proof adds

  • Patterns extracted from actual source code via AST
  • Every class, method, and dependency — mapped deterministically
  • 391 patterns from pydantic-ai + pydantic, validated by structure
  • User feedback loop: report success or failure, the network learns

Proof is a hosted MCP server that gives your AI coding assistant access to validated patterns from pydantic-ai's actual source code. Not docs. Not blog posts. The code itself.

391 patterns·1,251 nodes·6,946 relationships·1 endpoint

How it works

class Agent: def __init__(self): self.tools = [] @tool def search(q: str): return ... class Agent: def __init__(self): self.tools = [] @tool def search(q: str): return ... class Agent: def __init__(self): self.tools = [] @tool def search(q: str): return ... class Agent: def __init__(self): self.tools = [] @tool def search(q: str): return ... class Agent: def __init__(self): self.tools = [] @tool def search(q: str): return ... class Agent: def __init__(self): self.tools = [] @tool def search(q: str): return ... class Agent: def __init__(self): self.tools = [] @tool def search(q: str): return ... class Agent: def __init__(self): self.tools = [] @tool def search(q: str): return ... class Agent: def __init__(self): self.tools = [] @tool def search(q: str): return ... class Agent: def __init__(self): self.tools = [] @tool def search(q: str): return ...

Step 1 — Ingest

We analyze library source code to map every class, method, and dependency into a deterministic knowledge graph.

Step 2 — Extract

391 patterns are identified — constructor signatures, dependency chains, example code. Each starts at confidence 0.5 (code-derived).

Step 3 — Assemble

Your AI assistant queries via MCP. It searches, explains, and assembles working code from validated patterns. You report what worked. The network learns.

The 5 Tools

catalog()

Lists all libraries and pattern counts.

Example
catalog() → pydantic-ai: 391 patterns
When to use
"What do you know about?"
search(query)

Finds patterns by keyword.

Example
search("Agent") → 12 results, ranked
When to use
"Find patterns for X"
explain(symbol)

Deep-dive into a class or function — methods, dependencies, gotchas.

Example
explain("code.agent.Agent")
When to use
"Tell me about Agent"
build_agent(description)

Generates working agent code from validated patterns.

Example
build_agent("agent with tools and streaming")
When to use
"Build me an agent that..."
report(about, worked, details)

Reports whether the generated code worked. Drives confidence scores.

Example
report("agent_init", true, "ran without errors")
When to use
"After you test the code"
0
Patterns extracted
0
Knowledge nodes
0
Relationships mapped
0/10
Looking for first 10 testers

Add Proof to your AI assistant

The easiest way to connect. Click your client to see the 1-step method.

Run this in your terminal:
1-step setup
claude mcp add --transport http proof https://mcp.aigentys.com
That's it. Start a new Claude session and ask: "List the available MCP tools."
Don't see your client? Advanced setup for:

What's early

Confidence scores start at 0.5 (code-derived). They increase when users report success, decrease when they report failure.

Cross-library intelligence is in development. We are expanding the knowledge graph to support a wider range of AI frameworks and SDKs.

The network is new. Your feedback is what makes it smart.

Who this is for

  • You use pydantic-ai (or plan to)
  • You use Cursor, Claude, VS Code, or another MCP-compatible AI assistant
  • You're tired of generated code that calls APIs that don't exist
  • You want to be among the first 10 testers and shape the network

Who this is not for

  • You want a full production deployment right now (this is Phase 1.7)
  • You need patterns for other specific AI frameworks or SDKs (currently expanding support)
  • You want a GUI or dashboard (this is an MCP server — it plugs into your existing tools)