Integrations

This page lists the integration modules exported by @agentguard-run/spend 0.16.0. Each runs in the application process. Provider traffic continues directly from the application to the selected provider.

Exported TypeScript surface

OpenAI and OpenAI-compatible

@agentguard-run/spend/frameworks/openai

Wraps chat.completions.create before dispatch.

code →

Anthropic

@agentguard-run/spend/bindings/anthropic

Wraps Messages create and stream calls, then settles reported usage.

code →

AWS Bedrock

@agentguard-run/spend/bindings/bedrock

Gates InvokeModel and InvokeModelWithResponseStream commands.

code →

OpenRouter

@agentguard-run/spend/frameworks/openrouter

Wrap an OpenAI-compatible client or the chat-completions fetch endpoint.

code →

Vercel AI SDK

@agentguard-run/spend/frameworks/vercel-ai

Language-model middleware preflights generate and stream calls.

code →

LangChain.js

@agentguard-run/spend/frameworks/langchain

An awaited callback handler preflights LLM starts and settles usage at LLM end.

code →

Claude Code

@agentguard-run/spend/frameworks/claude-code

PreToolUse and PostToolUse hook handlers gate tools and append metadata records.

code →

Hermes

@agentguard-run/spend/frameworks/hermes

Pre-call, approval, and post-call hooks apply tool policy and record redacted metadata.

code →

Hermes Kanban

@agentguard-run/spend/frameworks/hermes-kanban

Task envelopes, capability ceilings, task caps, and parent-linked DAG receipts.

guide →

Provider bindings

The OpenAI-compatible binding wraps chat.completions.create. The Anthropic binding wraps messages.create and messages.stream. The Bedrock binding gates InvokeModel commands and passes unrelated commands through unchanged.

import { withSpendGuardOpenAI } from
  '@agentguard-run/spend/frameworks/openai';
import { withSpendGuardAnthropic } from
  '@agentguard-run/spend/bindings/anthropic';
import { withSpendGuardBedrock } from
  '@agentguard-run/spend/bindings/bedrock';

const openai = withSpendGuardOpenAI(openaiClient, { policy, scope });
const anthropic = withSpendGuardAnthropic(anthropicClient, { policy, scope });
const bedrock = withSpendGuardBedrock(bedrockClient, { policy, scope });

LangChain.js: full example

import { ChatOpenAI } from '@langchain/openai';
import { createLangChainHandler } from
  '@agentguard-run/spend/frameworks/langchain';

const handler = createLangChainHandler({ policy, scope });
const llm = new ChatOpenAI({ model: 'gpt-4o', callbacks: [handler] });
await llm.invoke('hello');

OpenRouter

import { withSpendGuardOpenRouter } from
  '@agentguard-run/spend/frameworks/openrouter';

const guarded = withSpendGuardOpenRouter(openRouterClient, { policy, scope });

Vercel AI SDK

import { wrapLanguageModel } from 'ai';
import { agentguardAiSdkMiddleware } from
  '@agentguard-run/spend/frameworks/vercel-ai';

const governed = wrapLanguageModel({
  model,
  middleware: agentguardAiSdkMiddleware({ policy, scope }),
});

Claude Code and Hermes hooks

import { createClaudeCodeHooks } from
  '@agentguard-run/spend/frameworks/claude-code';
import { createHermesPlugin } from
  '@agentguard-run/spend/frameworks/hermes';

const claudeHooks = createClaudeCodeHooks({ policy, scope });
const hermesPlugin = createHermesPlugin({ policy, scope });

Storage and signing boundaries

The package root exports InMemorySpendStore, InMemoryDecisionLogStore, and NdjsonDecisionLogStore. The in-memory stores are process-local. NDJSON is the file-backed decision-log option and is intended for a single process, with a best-effort lock file.

Redis and Postgres adapter classes exist in the 0.16.0 source tree but are not present in the package export map. Imports through internal dist/adapters paths are therefore not a supported consumer surface in this version.

Signed decision records require config.signingKeys with Ed25519 private and public key bytes. The current TypeScript package does not export an AWS KMS or Vault delegated signer. Without signing keys, policy enforcement still runs but no signed decision record is appended.

AgentGuard is patent-pending.