agntshare 0.1.0

HTTPS for AI Agent State. Lightweight, cryptographically verified Pathway Tokens to eliminate context bloat.

Agntshare (agnt.sr)

HTTPS for AI Agent State.

Today, AI agents hand off tasks by stuffing 100k+ tokens of raw memory into prompt windows. It’s slow, expensive, and degrades context.

Agntshare replaces raw context with a cryptographically verified "coat-check ticket" (Pathway Token). Mint a token on Agent A, pass a tiny 28-byte string, and let Agent B resolve the exact state it needs.

🚀 Quickstart

Install the zero-dependency SDK:

pip install agentshare

Zero-Friction Local Handoff (No API Keys Required):

from agentshare import AgentShareClient

# Initialize in local sandbox mode
client = AgentShareClient(mode="local")

# Agent A (Research) - Mints token for a massive state payload
state_payload = {"summary": "Deep-dive analysis...", "data": [1, 2, 3, 4, 5]}
token = client.mint_pathway_token(payload=state_payload, framework="python")
print(f"Token: agnt.sr/{token}") 

# Agent B (Writer) - Resolves only the keys it needs
resolved = client.resolve_pathway_token(token, keys=["summary"])
print(resolved['state'])

🛠️ Framework Integrations

LangChain Integration

from agentshare.integrations import AgentShareCallbackHandler

# 1-Line RunnableConfig integration
handler = AgentShareCallbackHandler(ttl_seconds=3600)
chain.invoke(input_data, config={"callbacks": [handler]})

Universal MCP Server (Claude Desktop & Cursor)

# Run stdio FastMCP server
python -m agentshare.integrations.mcp_server

---

🔒 Security & Inviolable Laws

1. Law 1: The Server is Blind (Zero-Trust Transport): The backend server NEVER parses or inspects payload bytes. Uploads and downloads use direct S3/R2 presigned URLs.
2. Law 2: The Token is the Truth: Pathway Tokens carry mandatory SHA-256 integrity checksums, TTL expiration window, scope rules, and cryptographic signatures.
3. Law 3: Zero-Friction Adoption: Integrations defensively swallow callback exceptions to ensure 0% impact on production agent executions.

---

License

MIT © Deepak V & Agntshare Contributors