Comparison
vs AI SDK 7
An honest, capability-by-capability comparison of @deuz-sdk/core v1.4 and Vercel AI SDK 7 to help you choose.
Both @deuz-sdk/core and the Vercel AI SDK are serious, multi-provider TypeScript AI SDKs used to build chatbots and agents in production. They overlap heavily on the core surface — streaming text, structured output, tool calling — and diverge in a few deliberate, architectural ways. This page is an honest capability comparison, not a sales pitch: it's drawn from a verified, code-level benchmark of both SDKs (docs/benchmark-ai-sdk.md in this repo), comparing @deuz-sdk/core v1.4.0 against AI SDK 7 (ai@7.0.14, GA 2026-06-25). Where AI SDK 7 is ahead, we say so plainly.
Capability comparison
| Capability | @deuz-sdk/core v1.4 | AI SDK 7 | Notes |
|---|---|---|---|
| Provider-agnostic chat | First-class — 6 provider families (Anthropic, OpenAI, xAI, Google, Vertex, Yunwu) behind one canonical call | First-class — 24 first-party providers + ~30 community | AI SDK's provider roster is far larger; see provider breadth below. |
| Canonical stream owned by the SDK | First-class — every response normalizes to a StreamPart union before any consumer sees it, with a synchronous, never-throwing contract (streamChat returns immediately; failures surface as an error part) | First-class — streamText also normalizes provider SSE into its own delta stream, with an onError callback for failures | Both own their wire and never proxy raw provider bytes. The contracts differ: Deuz's is a strict sync-return/never-throw guarantee; AI SDK's is callback-based and somewhat looser. |
| Agent loop shape | Free-function — generateText/streamChat gain an agentic loop the moment you pass tools, no agent object to construct | Class-based — ToolLoopAgent with stopWhen, prepareStep, activeTools, typed runtimeContext/toolsContext, lifecycle callbacks | Deuz keeps the loop a free function; AI SDK provides a dedicated agent class as its first-class abstraction. Both reach a similar end state. |
| Sub-agent live stream forwarding | First-class — agentTool forwards a sub-agent's entire canonical stream into the parent's fullStream as agentPath-tagged sub-agent parts, at any nesting depth | AI SDK documents a subagent (agent-as-tool) pattern, but live-forwarding the sub-agent's internal stream into the parent isn't part of its documented subagent surface | See Sub-agents. |
| Sub-agent tool approval | First-class in server mode — the parent's approveToolCall is inherited at every nesting depth, so a gated tool inside a sub-agent is decided by the same policy as the top-level loop | Not part of AI SDK's subagent tool surface — its own docs note that toolApproval doesn't compose with subagents today | Deuz's client-mode (pendingApprovals) approval inside a sub-agent is a known 1.4 limitation, planned for 1.5 alongside durable resume. |
| Automatic layered compaction | First-class — compaction: 'auto' prunes tool results, then reasoning, then summarizes the oldest slice, triggered automatically at a context-fill threshold | AI SDK provides manual pruning via a pruneMessages helper inside prepareStep; automatic, threshold-triggered compaction isn't part of the core surface yet (server-side OpenAI compaction is an open AI SDK feature request) | See Context compaction. |
| Token/cost budget stop conditions | First-class — totalTokensExceed(n) / costExceeds(usd) are built-in stopWhen factories, OR-ed with maxSteps | AI SDK's stopWhen supports step-count and tool-call conditions; usage is tracked, but a built-in token- or cost-budget stop condition isn't documented | |
Per-step loop hooks (prepareStep / activeTools) | First-class — CommonCallOptions.prepareStep runs after compaction, every step; activeTools filters the wire-visible tool list | First-class — ToolLoopAgent's prepareStep/activeTools are a core part of its design | Near parity in shape; Deuz composes it with automatic compaction, AI SDK composes it with its agent class. |
| Durable / resumable execution | Planned for 1.5 — a SessionStore + checkpoint/resume design is drafted but not shipped in 1.4 | First-class — WorkflowAgent makes each tool call a durable step with automatic retry and needsApproval suspend/resume across restarts | AI SDK advantage today. It requires the Vercel Workflow DevKit runtime ('use workflow' directives, stream()-only, serializable context) — a real capability with a corresponding runtime dependency. |
| Terminal UI | Not provided | First-class — @ai-sdk/tui | AI SDK advantage. |
| Harness adapters | Not provided — Deuz aims to be an SDK a harness is built on, not a wrapper around existing harnesses | First-class — HarnessAgent wraps Claude Code / Codex / Pi-style harnesses | AI SDK advantage if you want to wrap an existing harness rather than build on the SDK directly. |
| OpenTelemetry | A tracer seam exists in Dependencies for injecting your own tracer, but core spans aren't emitted yet in 1.4 | First-class — @ai-sdk/otel with GenAI semantic conventions, an invoke_agent/chat/execute_tool span hierarchy, and DevTools | AI SDK advantage today. |
| Provider breadth | ~6 provider families, each with a pinned capability/quirk registry | 24 first-party + ~30 community providers (~48+), including speech vendors, plus a provider registry with string-based lookup | AI SDK advantage. |
| UI framework breadth | React only (useChat/useObject) | React/Next, Vue/Nuxt, Svelte, Angular, Expo, TanStack Start | AI SDK advantage. |
| Speech / transcription | Not provided | First-class — generateSpeech/transcribe, including experimental streaming STT | AI SDK advantage. |
| Bundle size / zero deps | 0 runtime dependencies, ~2 MB installed, dual ESM+CJS, edge-safety enforced by lint | zod is a required peer, ESM-only (no CJS), typical install ~11–12 MB | Deuz advantage. |
| Determinism / injected seams | Every ambient effect (clock, randomness, id generation, fetch, logging) flows through one injected Dependencies seam — golden-replay tests run with no real network | Time and randomness are ambient in some code paths | Deuz advantage for reproducible tests and audits. |
| Cost tracking / pricing | Built-in — a token-to-USD pricing table plus priceProvider/onUsage seams in core | Cost tracking is pushed to the AI Gateway rather than the core SDK | Deuz advantage if you want cost visibility without adopting a gateway. |
Where Deuz-SDK leads today
- Free-function tool loop — parallel tool execution, self-healing, and runaway guards without an agent class to construct.
- A provider-native canonical stream with a strict contract —
streamChatreturns synchronously and never throws; failures are always anerrorpart or a rejected promise. - Sub-agent live stream forwarding and approval inheritance —
agentTool'ssub-agentparts and inheritedapproveToolCallcompose in one loop body, at any depth. - Automatic, layered context compaction — prune-then-summarize, triggered by a context-fill threshold, with cache-stable history.
- Token and cost budget stop conditions —
totalTokensExceed/costExceedsbuilt intostopWhen. - Pure dependency injection — clock, randomness, id generation, and fetch are all seamed, making agent behavior deterministic and replayable in tests.
- A zero-runtime-dependency, edge-safe-by-lint core — dual ESM+CJS, runs anywhere
fetchruns.
Where AI SDK 7 leads today
WorkflowAgent— durable, crash-resumable agent execution with automatic retry and suspend/resume, built on the Vercel Workflow DevKit runtime.@ai-sdk/tui— a ready-made terminal UI for CLI agents.HarnessAgent— wrappers for existing coding-agent harnesses (Claude Code, Codex, Pi-style tools).@ai-sdk/otel— first-class OpenTelemetry export with GenAI semantic conventions and DevTools.- Provider and UI-framework breadth — far more first-party/community providers, and support for React, Vue, Svelte, Angular, and Expo, not just React.
- Speech and transcription —
generateSpeech/transcribeare built in, including experimental streaming STT.
Choosing
Pick @deuz-sdk/core if you want a provider-native, deterministic, edge-safe foundation for agents or CLI tools where you want to own the loop, the stream, and the cost accounting directly, and you're comfortable with a narrower provider and UI-framework list today. Pick AI SDK 7 if you want the broadest provider and framework ecosystem, durable workflow execution out of the box, a terminal UI, or harness wrappers, and you're fine adopting its zod peer and (for WorkflowAgent) the Vercel Workflow DevKit runtime.
Durable, checkpoint/resume execution is on the Deuz roadmap for 1.5 — a storage-seamed design (no required vendor runtime), not yet shipped. If durable execution is a hard requirement today, AI SDK 7's WorkflowAgent is the more mature option.
See also
- Sub-agents —
agentTool, live stream forwarding, and approval inheritance. - Context compaction — the automatic, layered compaction policy.
- Tools —
prepareStep,activeTools, and budget stop conditions. - Migrating from Vercel AI SDK — API-level mapping between the two SDKs.