Your agent keeps its reasoning and workflow. PALISEC provides the authorization and protected-execution layer, the same way through every interface. Independent agents never need the native PALISEC Agent.
Your agent (any framework, any language) │ intent: to · value · data · gas limit · request ID ▼ TypeScript SDK │ Python SDK │ REST API │ MCP server ← authenticated integration principal (scoped API key) │ ▼ PALISEC authorization: deterministic engine · atomic reservation · replay · approvals │ └─ webhooks: signed events to your systems ▼ ALLOW → your agent sends exactly that transaction · DENY → nothing is sent · REVIEW → a human signs first
@palisec/sdk with viem or ethers. sendTransaction authorizes and sends only on ALLOW, refuses a wallet on the wrong chain or account.
palisec for async Python, standard library only, with a web3.py sender. Same semantics as the TypeScript SDK.
OpenAPI 3.1. Idempotent authorizations, dry runs, operation status, the agent's policy and audit trail.
Streamable HTTP at /api/mcp with the official MCP SDK. Tools for evaluation, proposals, status, explanations and policy validation.
Signed events for decisions, approvals, pauses, stops and security findings, with stable event IDs for deduplication.
createPalisecTools in the TypeScript SDK gives every framework the same tested tools: evaluate, send (only on ALLOW), read an operation, explain a reason code. There is no framework-specific security logic, and MCP-capable frameworks can use the MCP server instead.
The core SDKs stay framework-independent on purpose.
tools.ts
import { createPalisec, createPalisecTools } from "@palisec/sdk";
// Framework-neutral tools: name, description, JSON Schema, execute.
// The framework decides what to try; PALISEC decides whether it may be sent.
const tools = createPalisecTools(createPalisec({ apiKey: process.env.PALISEC_API_KEY! }), { wallet });
// OpenAI Agents SDK
new Agent({ name: "trader", tools: tools.map((t) => tool({ ...t, strict: false })) });
// Vercel AI SDK
Object.fromEntries(tools.map((t) => [t.name, aiTool({ description: t.description, inputSchema: jsonSchema(t.parameters), execute: t.execute })]));
// LangGraph / LangChain
tools.map((t) => new DynamicStructuredTool({ name: t.name, description: t.description, schema: t.parameters, func: t.execute }));Each integration is a registered agent bound to one protected account and policy. Its API keys are shown once, stored only as SHA-256 hashes, and can be narrowed to the scopes it needs, rotated atomically or revoked instantly. A name is not an identity: the key is.
An integration can never