AI SDK Middleware

Reusable middleware utilities for the Vercel AI SDK v7 (LanguageModelV4), plus an evlog/ai LanguageModelV4 port.

Support matrix

Aligned with the official AI SDK layering and evlog AI integration:

| Layer | Spec | |-------|------| | AI SDK v6 | Primarily LanguageModelV3 / LanguageModelV3Middleware | | AI SDK v7 (ai@7) | Core is LanguageModelV4 (specificationVersion: "v4") | | This package (core middlewares) | LanguageModelV4Middleware only | | @ai-sdk-tool/middleware/evlog wrap | LanguageModelV4 + LanguageModelV4Middleware | | Upstream evlog/ai wrap | LanguageModelV3 + LanguageModelV3Middleware only | | Peer ai (optional, for ./evlog) | >=7.0.0 <8.0.0 (LanguageModelV4 wrap — not AI SDK v6) | | Peer evlog (optional, for ./evlog types) | >=2.0.0 — full RequestLogger type | | Peer @ai-sdk/provider | ^4 (LanguageModelV4 types) | | Node.js | >=22 (required for AI SDK v7) |

Meaning: this package targets AI SDK v7 / LanguageModelV4 only. Upstream evlog/ai still peers ai >= 6.0.168 and wraps V3; use that for AI SDK v6.

./evlog feature matrix

Capture token usage, tool calls, model info, and streaming metrics into wide events. Requires AI SDK v7 (ai >= 7, Node.js 22+).

For tool execution timing, abort tracking, and auto embed capture, pass createEvlogIntegration(ai) to telemetry.integrations.

| Data | Source | Description | |------|--------|-------------| | tokens, model, provider, stream metrics | middleware (wrap / createAIMiddleware) | inputTokens, outputTokens, cache/reasoning tokens, msToFirstChunk, msToFinish, tokensPerSecond, … | | ai.tools[] | onToolExecutionEnd | Per-tool name, durationMs, success, error | | ai.totalDurationMs | onStart → onEnd | Wall time from generation start to completion | | ai.embedding | onEmbedEnd or captureEmbed() | Embeddings | | ai.finishReason: 'abort' | onAbort | Aborted stream | | ai.error | onAbort / onError | Abort reason or unrecoverable error |

Integration still implements the older v6 hook names (onToolCallFinish, onFinish) for source parity with upstream, but this package only supports AI SDK v7.

Installation

pnpm add @ai-sdk-tool/middleware
# for ./evlog:
pnpm add ai@^7

Exports

• @ai-sdk-tool/middleware
• @ai-sdk-tool/middleware/disk-cache
• @ai-sdk-tool/middleware/reasoning-parser
• @ai-sdk-tool/middleware/evlog — LanguageModelV4 port of evlog/ai (optional peer: ai)

Included middleware

All middlewares implement LanguageModelV4Middleware and work with wrapLanguageModel from ai@7:

• createDiskCacheMiddleware: Disk-based response cache for generate and stream
• defaultSystemPromptMiddleware: Inject or merge system prompts
• extractReasoningMiddleware: Extract XML-tagged reasoning into reasoning parts
• createAIMiddleware / createAILogger / createEvlogIntegration (./evlog): wide-event AI observability (log.set({ ai }))

evlog example (v7)

import { generateText } from "ai";
import {
  createAILogger,
  createEvlogIntegration,
} from "@ai-sdk-tool/middleware/evlog";

const ai = createAILogger(log); // full evlog RequestLogger

const result = await generateText({
  model: ai.wrap(yourLanguageModelV4), // or gateway model id string
  prompt: "hello",
  telemetry: {
    integrations: [createEvlogIntegration(ai)],
  },
});

Core middleware example

import { wrapLanguageModel } from "ai";
import { createDiskCacheMiddleware } from "@ai-sdk-tool/middleware/disk-cache";
import { defaultSystemPromptMiddleware } from "@ai-sdk-tool/middleware";

const model = wrapLanguageModel({
  model: yourLanguageModelV4,
  middleware: [
    defaultSystemPromptMiddleware({ systemPrompt: "You are helpful." }),
    createDiskCacheMiddleware({ cacheDir: ".ai-cache" }),
  ],
});

Development

pnpm install
pnpm run typecheck
pnpm run test
pnpm run build

Release flow

• Add a patch/minor/major changeset in .changeset/*.md
• Merge changes to main via a pull request, then let Release Changeset open/update the version PR
• Merging the version PR publishes the package to npm automatically