Where It Fits
Where does AI SDK belong in the stack?
AI SDK's primary role: TypeScript Agent & AI App Toolkit. It is Core across Build, supports Model, Data, and Operate, and touches Evaluate.
Model
- Model APIs: Touches
- Model gateway: Supports
Build
- Orchestration: Core
- Tool use: Core
Data
- Retrieval: Supports
Evaluate
- Evaluation: Touches
Operate
- Observability: Supports
- Cloud deployment: Touches
- Guardrails: Touches
AI SDK beside the rest of its stack
Every product we cover on AI Application Platform, on the layers it owns, placed by what its practitioners report. Pick any of them to compare.
Signals from
Show products that
Activating a product opens its read here. Activating the selected product opens its hub.
Each product sits on the layers it owns, placed left to right by how far its positive sources lead its negative ones. A disc's area is its sources in this view and its tone is its band. Products with fewer than 20 sources from 15 authors have no net and sit at the right, under Unscored. The teal line joins AI SDK across its layers.
Read the field as a table
| Product | Core on | Supports | Overall Signal | Net | Sources | Positive | Mixed | Negative |
|---|---|---|---|---|---|---|---|---|
| Claude Agent SDK | Orchestration | Tool use, Cloud deployment, Guardrails | Strong positive | +52 | 22 | 13 | 8 | 1 |
| Braintrust | Evaluation, Observability | Model gateway, Tool use | Strong positive | +50 | 34 | 20 | 11 | 3 |
| Langfuse | Evaluation, Observability | None | Positive | +46 | 26 | 12 | 14 | 0 |
| Jev | Supporting only | Model APIs, Model gateway, Tool use, Retrieval, Evaluation, Guardrails | Positive | +40 | 25 | 12 | 11 | 2 |
| Datadog Agent Observability | Evaluation, Observability | Guardrails | Positive | +32 | 26 | 9 | 16 | 1 |
| Arize | Evaluation, Observability | Guardrails | Positive | +26 | 27 | 11 | 12 | 4 |
| Claude API | Model APIs, Tool use | Orchestration, Retrieval, Cloud deployment | Positive | +25 | 52 | 22 | 21 | 9 |
| Amazon Bedrock | Model APIs, Cloud deployment | Model gateway, Orchestration, Tool use, Retrieval, Evaluation, Observability, Guardrails | Mixed | +9 | 57 | 25 | 12 | 20 |
| LiteLLM | Model gateway | Tool use, Observability, Guardrails | Mixed | +8 | 37 | 10 | 21 | 6 |
| OpenAI API | Model APIs, Tool use | Orchestration, Retrieval, Observability, Cloud deployment, Guardrails | Mixed | +3 | 65 | 22 | 23 | 20 |
| Gemini API | Model APIs, Tool use | Orchestration, Retrieval, Cloud deployment, Guardrails | Mixed | 0 | 74 | 24 | 25 | 25 |
| pgvector | Retrieval | Cloud deployment | Negative | −29 | 24 | 5 | 8 | 11 |
| Mastra | Orchestration | Model gateway, Tool use, Retrieval, Evaluation, Observability, Cloud deployment, Guardrails | Negative | −36 | 25 | 4 | 8 | 13 |
| Qdrant | Retrieval | Cloud deployment | Negative | −40 | 41 | 8 | 10 | 23 |
| Pinecone | Retrieval | Tool use, Cloud deployment | Strong negative | −63 | 33 | 3 | 6 | 24 |
| LangSmith | Evaluation, Observability | Model gateway, Orchestration, Tool use, Cloud deployment | Below minimum | none | 19 | 9 | 7 | 3 |
| AI SDK | Orchestration, Tool use | Model gateway, Retrieval, Observability | Below minimum | none | 18 | 8 | 9 | 1 |
| LangGraph | Orchestration | Tool use, Cloud deployment | Below minimum | none | 18 | 10 | 5 | 3 |
| Portkey | Model gateway | Observability, Guardrails | Below minimum | none | 18 | 11 | 4 | 3 |
| Fireworks AI | Cloud deployment | Model APIs, Model gateway, Retrieval | Below minimum | none | 16 | 6 | 5 | 5 |
| Agent Platform | Model APIs, Cloud deployment | Model gateway, Orchestration, Tool use, Retrieval, Evaluation, Observability, Guardrails | Below minimum | none | 13 | 7 | 1 | 5 |
| OpenRouter | Model gateway | Model APIs, Observability, Guardrails | Below minimum | none | 11 | 4 | 4 | 3 |
| Microsoft Foundry | Model APIs, Cloud deployment | Model gateway, Orchestration, Tool use, Retrieval, Evaluation, Observability, Guardrails | Below minimum | none | 8 | 1 | 2 | 5 |
| OpenAI Agents SDK | Orchestration | Tool use, Observability, Guardrails | Below minimum | none | 7 | 3 | 2 | 2 |
| CrewAI | Orchestration | Tool use, Retrieval, Observability, Cloud deployment, Guardrails | Below minimum | none | 4 | 1 | 3 | 0 |
| Together AI | Cloud deployment | Model APIs | Below minimum | none | 1 | 0 | 0 | 1 |
| DeepSeek API | Model APIs | Tool use, Retrieval | Below minimum | none | 0 | 0 | 0 | 0 |
Each author counts once.
Stack Footprint
What is AI SDK used for, and by whom?
Agent Frameworks, also listed under AI engineering and developer tools.
Model
Model APIs: Touches
Model APIs, touches. Provider packages call OpenAI, Anthropic, Google, Amazon Bedrock, Azure, and 40 more first-party integrations. The AI SDK supplies no models. Each provider bills the tokens.
Model gateway: Supports
Model gateway, supports. Vercel AI Gateway resolves plain model-id strings as the default global provider. The separate paid service adds budgets and fallbacks at zero token markup. The SDK itself switches providers in code.
Build
Orchestration: Core
Orchestration, core. generateText and streamText run multi-step tool loops. ToolLoopAgent packages agents, while WorkflowAgent adds durability through Workflow DevKit. HarnessAgent runs Claude Code, Codex, or Pi through the same interface.
Tool use: Core
Tool use, core. Tools accept Zod or JSON Schema inputs and per-tool context. The SDK supports provider-executed tools and an MCP client with MCP Apps. Approvals pause tools for a human or an approval function.
Data
Retrieval: Supports
Retrieval, supports. embed and embedMany call embedding models across providers, while rerank calls reranking models. A vector store handles retrieval because the SDK stores and searches no documents.
Evaluate
Evaluation: Touches
Evaluation, touches. ai/test supplies model helpers, and response objects record every step. Teams such as Xata use them in Vitest evals. An evaluation tool owns datasets and scoring.
Operate
Observability: Supports
Observability, supports. AI SDK 7 registers OpenTelemetry once for all calls. It emits Node.js tracing-channel events and per-step timing. Datadog, Langfuse, Braintrust, Sentry, and LangSmith integrations provide trace storage and analysis.
Cloud deployment: Touches
Cloud deployment, touches. The SDK runs on Node.js 22 or later, on Vercel or another cloud. The provider or AI Gateway controls model capacity and quotas. Host function timeouts limit long runs.
Guardrails: Touches
Guardrails, touches. AI SDK language model middleware intercepts and modifies model calls and provides a place to add guardrails. The SDK does not supply a built-in prompt or response screening service.
Used for AI feature development, agent development, and prototyping, by software engineers, AI engineers, and product teams.
- Business Area
- Product & Engineering
- Function
- Software Development
- Operating Stack
- AI Application Platform
- Primary Role
- TypeScript Agent & AI App Toolkit
How to read this map. Each layer lists the products that structurally own it, so a product that only touches a layer isn't listed among its owners. AI SDK is underlined in teal where it is Core, and its rows sit a shade lighter; where it supports or touches a layer, that row names it first.
Model
AI SDK: supports on 1 of 2 layers; touches on 1 of 2 layers
Model APIs
Hosted models an application calls for generation, reasoning, and vision, under the provider's rate limits, pricing, and data terms.
- AI SDK touchesAdjacentProvider packages call OpenAI, Anthropic, Google, Amazon Bedrock, Azure, and 40 more first-party integrations. The AI SDK supplies no models. Each provider bills the tokens.
- Amazon BedrockAI Model APIAmazon BedrockAmazon Web Services+9 MixedAWS service for building on models from Amazon, Anthropic, OpenAI, DeepSeek, and others. AgentCore, Knowledge Bases, and Guardrails add agents, retrieval, and filters. Usage bills per token on AWS.Owns this layerModel APIsOpen product page
- Claude APIAI Model APIClaude APIAnthropic+25 PositiveAnthropic's developer platform for building AI features and agents on Claude models: the Messages API, tool use, prompt caching, batch processing, and the Agent SDK.Owns this layerModel APIsOpen product page
- DeepSeek APIAI Model APIDeepSeek APIDeepSeek+57 Strong positiveDeepSeek's first-party API serves deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro with a 1M-token context, in OpenAI- and Anthropic-compatible formats. It bills per million tokens from a prepaid balance, with cheap cache hits and half-price off-peak hours, and stores data in the People's Republic of China.Owns this layerModel APIsOpen product page
- Gemini APIAI Model APIGemini APIGoogle0 MixedGoogle's API for Gemini features and agents, reached with an AI Studio key or the Gemini Enterprise Agent Platform. It supports function calling, built-in tools, context caching, batch, Flex, and Priority inference.Owns this layerModel APIsOpen product page
- Gemini Enterprise Agent PlatformAI Model APIGemini Enterprise Agent PlatformGoogle−27 NegativeGoogle Cloud platform, formerly Vertex AI, for building on Gemini, Claude and open models from Model Garden. The Agent Development Kit, Agent Runtime and RAG Engine add agents and retrieval. Usage bills per token through Google Cloud.Owns this layerModel APIsOpen product page
- Microsoft FoundryAI Model APIMicrosoft FoundryMicrosoft−26 NegativeAzure platform, formerly Azure AI Foundry, for building on OpenAI, Anthropic, Meta, DeepSeek and Microsoft models. Foundry Agent Service, evaluations and model router add agents and routing. Usage bills per token or per provisioned unit on Azure.Owns this layerModel APIsOpen product page
- OpenAI APIAI Model APIOpenAI APIOpenAI+3 MixedOpenAI's platform for teams building AI features and agents on GPT models. The Responses API supports function calling and remote MCP servers. Prompt caching, Batch, and Flex reduce usage costs, and the Agents SDK runs agents.Owns this layerModel APIsOpen product page
Model gateways & routing
One interface in front of several models or providers that routes requests, fails over, and enforces spend and rate limits.
- AI SDK supportsSupportingVercel AI Gateway resolves plain model-id strings as the default global provider. The separate paid service adds budgets and fallbacks at zero token markup. The SDK itself switches providers in code.
- LiteLLMAI GatewayLiteLLMBerriAI+8 MixedBerriAI's self-hosted gateway and Python SDK expose 100+ model providers through one OpenAI-format API. The MIT-licensed core includes virtual keys, budgets, rate limits, fallbacks, and spend tracking. Enterprise adds SSO and audit logs, plus support.Owns this layerModel gateways & routingOpen product page
- OpenRouterAI GatewayOpenRouterOpenRouter+10 MixedOpenRouter's hosted gateway routes one OpenAI-compatible API to 500+ models from 80+ providers, with provider fallback, per-key credit limits, workspaces, guardrails, and data-policy routing. Tokens bill at provider list price from prepaid credits, with a fee on credit purchases. Stripe agreed to acquire OpenRouter on August 19, 2026.Owns this layerModel gateways & routingOpen product page
- PortkeyAI GatewayPortkeyPalo Alto Networks0 MixedPortkey's gateway puts one OpenAI-compatible API, fallbacks, load balancing, request logs, and guardrails in front of a team's own model API keys, as a hosted service, a hybrid data plane in your VPC, or an MIT-licensed open-source gateway. Palo Alto Networks completed its purchase on May 29, 2026 and sells it as Prisma AIRS AI Gateway.Owns this layerModel gateways & routingOpen product page
- Vercel AI Gateway
Build
AI SDK: core on 2 of 2 layers
Orchestration & agent frameworks
Code that composes model calls, tools, memory, and control flow into multi-step workflows and agents.
- AI SDK
- Claude Agent SDKAgent FrameworkClaude Agent SDKAnthropic+52 Strong positiveAnthropic's SDK, formerly the Claude Code SDK, runs the Claude Code binary inside your own service, with built-in tools, subagents, hooks, MCP, permissions, and sessions. It calls Claude models only and is billed at Claude API token prices or against Max and Team plan credits.Owns this layerOrchestration & agent frameworksOpen product page
- CrewAIAgent FrameworkCrewAICrewAI−7 MixedCrewAI's MIT-licensed Python framework builds crews of role-based agents inside event-driven flows, on models from any vendor. CrewAI AMP deploys, traces, and governs them, free for 50 executions a month and custom-priced on Enterprise.Owns this layerOrchestration & agent frameworksOpen product page
- LangGraphAgent FrameworkLangGraphLangChain+25 PositiveLangChain's MIT-licensed runtime builds agents as graphs with shared state in Python and JavaScript, with or without LangChain. It adds checkpoints, human interrupts, and streaming. LangSmith Deployment offers paid hosting.Owns this layerOrchestration & agent frameworksOpen product page
- MastraAgent FrameworkMastraMastra−36 NegativeMastra is a TypeScript framework for agents, typed workflows, memory, and evals, Apache 2.0 outside its ee/ directories. Mastra Platform hosts and traces them: Starter is free with monthly allowances, Teams is US$250 a month, and Enterprise is custom.Owns this layerOrchestration & agent frameworksOpen product page
- OpenAI Agents SDKAgent FrameworkOpenAI Agents SDKOpenAI−17 MixedOpenAI's MIT-licensed SDK runs agents with tools, handoffs, guardrails, sessions, approvals, MCP tools, and tracing inside your own service. It calls OpenAI models through the Responses API by default, reaches other models through beta adapters, and bills at OpenAI API token prices.Owns this layerOrchestration & agent frameworksOpen product page
Tool use & integrations
Letting a model call functions, APIs, and MCP servers, and act on what they return.
- AI SDK
- Claude APIAI Model APIClaude APIAnthropic+25 PositiveAnthropic's developer platform for building AI features and agents on Claude models: the Messages API, tool use, prompt caching, batch processing, and the Agent SDK.Owns this layerTool use & integrationsOpen product page
- Gemini APIAI Model APIGemini APIGoogle0 MixedGoogle's API for Gemini features and agents, reached with an AI Studio key or the Gemini Enterprise Agent Platform. It supports function calling, built-in tools, context caching, batch, Flex, and Priority inference.Owns this layerTool use & integrationsOpen product page
- OpenAI APIAI Model APIOpenAI APIOpenAI+3 MixedOpenAI's platform for teams building AI features and agents on GPT models. The Responses API supports function calling and remote MCP servers. Prompt caching, Batch, and Flex reduce usage costs, and the Agents SDK runs agents.Owns this layerTool use & integrationsOpen product page
Data
AI SDK: supports on 1 of 1 layer
Retrieval & grounding
Embedding, indexing, and retrieving the documents and records an application grounds its answers in.
- AI SDK supportsSupportingembed and embedMany call embedding models across providers, while rerank calls reranking models. A vector store handles retrieval because the SDK stores and searches no documents.
- pgvectorVector DatabasepgvectorThe pgvector project−29 NegativePostgres extension that stores vectors with the rest of your data and searches them exactly or with HNSW and IVFFlat indexes. Free under the PostgreSQL License, self-installed or on managed hosts such as Neon, Supabase, Cloud SQL, Amazon RDS and Azure.Owns this layerRetrieval & groundingOpen product page
- PineconeVector DatabasePineconePinecone Systems−63 Strong negativeManaged vector database serving full-text, semantic, sparse and hybrid search from one index, with hosted embedding and reranking and namespaces per tenant. Runs on AWS, GCP and Azure, or in your own cloud with Bring Your Own Cloud on Enterprise.Owns this layerRetrieval & groundingOpen product page
- QdrantVector DatabaseQdrantQdrant Solutions GmbH−40 NegativeVector database for dense, sparse and multivector search with JSON payload filters, hybrid search and built-in quantization. Self-hosted under the Apache License 2.0, or Qdrant Cloud on AWS, GCP and Azure with a free tier and usage-based Standard billing.Owns this layerRetrieval & groundingOpen product page
- Voyage AI
Evaluate
AI SDK: touches on 1 of 1 layer
Evaluation & testing
Scoring model and agent output against datasets and graders, so a prompt or model change is measured before it ships.
- AI SDK touchesAdjacentai/test supplies model helpers, and response objects record every step. Teams such as Xata use them in Vitest evals. An evaluation tool owns datasets and scoring.
- ArizeAI ObservabilityArizeArize AI+26 PositiveArize AX and Phoenix trace and evaluate LLM apps and agents over OpenTelemetry. Managed AX adds online evals and monitors, plus Signal. Phoenix is free to self-host under the Elastic License 2.0. Arize has been part of Dynatrace since October 2026.Owns this layerEvaluation & testingOpen product page
- BraintrustAI ObservabilityBraintrustBraintrust+50 Strong positiveBraintrust evaluates and traces AI apps and agents. Dataset experiments and scorers test prompt and model changes, plus code, before release. Production tracing and online scoring, plus Topics, track results afterward. It runs in Braintrust's US or EU cloud, or, on Enterprise, with a data plane in your own cloud.Owns this layerEvaluation & testingOpen product page
- Datadog Agent ObservabilityAI ObservabilityDatadog Agent ObservabilityDatadog+32 PositiveDatadog Agent Observability, sold as LLM Observability until 2026, traces LLM apps and agents and evaluates their output. It shares one trace ID with Datadog APM, infrastructure, and user monitoring. It bills only LLM spans, one per model call, and can be bought without other Datadog products.Owns this layerEvaluation & testingOpen product page
- LangfuseAI ObservabilityLangfuseClickHouse+46 PositiveLangfuse traces and evaluates LLM apps and agents. It versions prompts on Langfuse Cloud or self-hosted under the MIT license. ClickHouse acquired it in January 2026.Owns this layerEvaluation & testingOpen product page
- LangSmithAI ObservabilityLangSmithLangChain+17 MixedLangChain's platform for teams tracing and evaluating LLM applications and agents. It also deploys agents on LangChain's cloud or your infrastructure, with or without the LangChain frameworks.Owns this layerEvaluation & testingOpen product page
Operate
AI SDK: supports on 1 of 3 layers; touches on 2 of 3 layers
Observability & tracing
Tracing production requests across model calls and agent steps, with the cost, latency, and errors of each.
- AI SDK supportsSupportingAI SDK 7 registers OpenTelemetry once for all calls. It emits Node.js tracing-channel events and per-step timing. Datadog, Langfuse, Braintrust, Sentry, and LangSmith integrations provide trace storage and analysis.
- ArizeAI ObservabilityArizeArize AI+26 PositiveArize AX and Phoenix trace and evaluate LLM apps and agents over OpenTelemetry. Managed AX adds online evals and monitors, plus Signal. Phoenix is free to self-host under the Elastic License 2.0. Arize has been part of Dynatrace since October 2026.Owns this layerObservability & tracingOpen product page
- BraintrustAI ObservabilityBraintrustBraintrust+50 Strong positiveBraintrust evaluates and traces AI apps and agents. Dataset experiments and scorers test prompt and model changes, plus code, before release. Production tracing and online scoring, plus Topics, track results afterward. It runs in Braintrust's US or EU cloud, or, on Enterprise, with a data plane in your own cloud.Owns this layerObservability & tracingOpen product page
- Datadog Agent ObservabilityAI ObservabilityDatadog Agent ObservabilityDatadog+32 PositiveDatadog Agent Observability, sold as LLM Observability until 2026, traces LLM apps and agents and evaluates their output. It shares one trace ID with Datadog APM, infrastructure, and user monitoring. It bills only LLM spans, one per model call, and can be bought without other Datadog products.Owns this layerObservability & tracingOpen product page
- LangfuseAI ObservabilityLangfuseClickHouse+46 PositiveLangfuse traces and evaluates LLM apps and agents. It versions prompts on Langfuse Cloud or self-hosted under the MIT license. ClickHouse acquired it in January 2026.Owns this layerObservability & tracingOpen product page
- LangSmithAI ObservabilityLangSmithLangChain+17 MixedLangChain's platform for teams tracing and evaluating LLM applications and agents. It also deploys agents on LangChain's cloud or your infrastructure, with or without the LangChain frameworks.Owns this layerObservability & tracingOpen product page
Cloud deployment & capacity
Running model workloads inside a cloud account or region with the throughput, data residency, and billing the business requires.
- AI SDK touchesAdjacentThe SDK runs on Node.js 22 or later, on Vercel or another cloud. The provider or AI Gateway controls model capacity and quotas. Host function timeouts limit long runs.
- Amazon BedrockAI Model APIAmazon BedrockAmazon Web Services+9 MixedAWS service for building on models from Amazon, Anthropic, OpenAI, DeepSeek, and others. AgentCore, Knowledge Bases, and Guardrails add agents, retrieval, and filters. Usage bills per token on AWS.Owns this layerCloud deployment & capacityOpen product page
- Fireworks AIInference HostingFireworks AIFireworks AI−17 MixedFireworks AI serves open models through Serverless Inference, multi-tenant inference on Fireworks-managed infrastructure billed per token, and through on-demand deployments on dedicated GPUs billed per GPU second. It offers supervised and reinforcement fine-tuning and serves fine-tuned models for the same price as base models.Owns this layerCloud deployment & capacityOpen product page
- Gemini Enterprise Agent PlatformAI Model APIGemini Enterprise Agent PlatformGoogle−27 NegativeGoogle Cloud platform, formerly Vertex AI, for building on Gemini, Claude and open models from Model Garden. The Agent Development Kit, Agent Runtime and RAG Engine add agents and retrieval. Usage bills per token through Google Cloud.Owns this layerCloud deployment & capacityOpen product page
- Microsoft FoundryAI Model APIMicrosoft FoundryMicrosoft−26 NegativeAzure platform, formerly Azure AI Foundry, for building on OpenAI, Anthropic, Meta, DeepSeek and Microsoft models. Foundry Agent Service, evaluations and model router add agents and routing. Usage bills per token or per provisioned unit on Azure.Owns this layerCloud deployment & capacityOpen product page
- Together AIInference HostingTogether AITogether AI−49 NegativeTogether AI serves open models through a serverless per-token API compatible with the OpenAI SDKs, provisioned throughput with a monthly success commitment, dedicated endpoints billed by the minute, fine-tuning and multi-node GPU clusters. Accounts are fully prepaid, with a minimum US$15 credit purchase.Owns this layerCloud deployment & capacityOpen product page
- Microsoft Foundry
- Groq
Guardrails & AI security
Screening prompts, responses, and agent actions for prompt injection, sensitive data, and policy violations, and blocking or logging what fails.
- AI SDK touchesAdjacentAI SDK language model middleware intercepts and modifies model calls and provides a place to add guardrails. The SDK does not supply a built-in prompt or response screening service.
- Prisma AIRS
- NeMo Guardrails
