@goderash/sdk

Audit any AI agent with one import. Tamper-evident, hash-chained, regulator-ready.

Goderash is the audit & governance fabric for regulated AI agents. One decorator turns every tool call, LLM call, and policy decision into a typed, immutable, SHA-256 hash-chained event in a per-tenant ledger. From there: signed evidence packs for SOC 2, HIPAA, FFIEC, FINRA, and SEC Rule 17a-4 in one click.

• 🌐 Live demo: ai.goderash.com
• 📚 Repo: github.com/goderash/goderash
• 🐍 Python: goderash-sdk

Install

npm i @goderash/sdk
# or
pnpm add @goderash/sdk
# or
yarn add @goderash/sdk

60-second quickstart

import { GoderashClient, wrapTool } from '@goderash/sdk'

const goderash = new GoderashClient({
  apiKey: process.env.GODERASH_API_KEY!,
  tenant: 'acme-finance',
  agentId: 'portfolio-advisor-v1',
})

const transferMoney = wrapTool(
  goderash,
  { name: 'transfer_money', category: 'action', confirmation: 'biometric' },
  async (src: string, dst: string, amount: number) => {
    // ... your logic ...
    return { status: 'queued' }
  },
)

const ctx = goderash.newContext()
await transferMoney('checking', 'savings', 100, { goderashContext: ctx })
// → emits: tool.invoked, tool.completed
// → hash-chained into the per-tenant ledger
// → ready to be included in your next SOC 2 / HIPAA evidence pack

What wrap* does

| Decorator | What it wraps | Events emitted | |---|---|---| | wrapTool | Any function the agent can call | tool.invoked, tool.completed / tool.failed, plus any contract.violated events | | wrapLLM | Your LLM call (OpenAI, Anthropic, etc.) | llm.call.started, llm.call.completed (with token usage) | | wrapAgent | The full agent turn boundary | agent.turn.started, agent.turn.completed (parent of every tool/LLM event in the turn) |

Every event carries tenant_id, agent_id, conversation_id, turn_id, parent_event_id, occurred_at, recorded_at, prev_hash, hash — the chain-of-custody auditors require.

Runtime guards

The same SDK ships the runtime safety stack we built for an AI banking agent in production:

import {
  GuardChain,
  FraudGuard,
  PermissionModeGate,
  VelocityLimiter,
  ConversationBudget,
  CancellationToken,
} from '@goderash/sdk/guards'

const guards = new GuardChain([
  new PermissionModeGate({ mode: 'STRICT' }),
  new FraudGuard(),
  new VelocityLimiter({
    rules: {
      transfer_money: [{ windowSec: 3600, maxCount: 5, label: '5/hour' }],
    },
  }),
  new ConversationBudget({ maxToolCalls: 20, maxTokens: 200_000 }),
])

const decision = await guards.evaluate(goderash, ctx, {
  tool: 'transfer_money',
  input: { amount: 500 },
})
if (decision.allow) {
  await transferMoney(...)
}

| Guard | What it stops | |---|---| | PermissionModeGate | Tools your current mode (PLAN / DEFAULT / AUTO / STRICT) doesn't authorize | | FraudGuard | Pattern-based attack signatures at the input boundary | | VelocityLimiter | Per-tool rate limit (Redis-backed in production) | | ConversationBudget | Runaway loops (max tool calls, max tokens per conversation) | | CancellationToken | Irreversible actions cancelled mid-flight |

Drop-in framework adapters

| Framework | Package | |---|---| | Anthropic Messages / Claude SDK | @goderash/adapter-claude-sdk | | OpenAI Assistants | @goderash/adapter-openai-assistants | | LangGraph.js / LangChain.js | @goderash/adapter-langgraph |

License

Apache-2.0