Mastra

Front-end AI ecosystem frameworks include LangChain.js/LangGraph.js (agent orchestration), RAGFlow.ts (RAG engineering), Mastra (Agent development), and LlamaIndex.ts (data indexing and retrieval layer).

Under the hood, Mastra uses the Vercel AI SDK's @ai-sdk/openai-compatible provider to communicate with custom endpoints, sending requests to /v1/chat/completions with the user agent ai-sdk/openai-compatible.

Mastra supports pointing to a custom OpenAI-compatible endpoint by setting the model field to an object containing id, url, and apiKey, with no additional provider package needed.

Setting the agent's errorProcessorDefaults to false or passing errorProcessors: [] in the generate call disables the outer error handler retries in Mastra.

Mastra's AI SDK retry layer honors the Retry-After response header when encountering a 429.

In Mastra, with modelSettings.maxRetries set to 5 and the default error handler retained, a single upstream 503 failure results in 18 requests and about 99 seconds of waiting.

Mastra has two layers of retry mechanisms: the AI SDK's modelSettings.maxRetries (default 0 in Mastra) and Mastra's default error handler (two additional retries, roughly three seconds apart); when both are enabled, failed requests multiply.

A Mastra agent fails immediately on a 401 error without retrying.

A Mastra agent's tool-calling loop silently stops after reaching the default step limit (5 steps), returning empty text with finishReason "tool-calls", without throwing an error or logging anything.

Setting api: "responses" in the Mastra model configuration causes the agent to instead POST to /v1/responses and use an input array in place of messages.

When apiKey is omitted in a Mastra custom model configuration, there is no fallback to reading environment variables, and requests go out without an Authorization header, resulting in a 401.

Mastra does not automatically complete the /v1 path for the url; the user must include /v1 in the url themselves, otherwise the request is sent to /chat/completions rather than /v1/chat/completions.

In Mastra's model id, everything before the first slash is stripped, and only what follows is sent to the endpoint as the model field.

When Mastra configures model using the object form, it does not issue a GET /v1/models probe and does not add extra parameters.

Mastra's agent.generate supports controlling the maximum number of tool-calling loop rounds via the maxSteps parameter.

The choice between Mastra and LangGraph should be based on the topological complexity of the agent flow rather than ecosystem preference: linear tasks are lighter with an integrated framework, while multi-branch flows that require rollback are better suited to a graph structure.

Mastra is a TypeScript-first agent framework that provides integrated capabilities such as workflows, tools, memory, and RAG.

LangGraph's graph state maps naturally to state update events in the event stream; Mastra's workflow steps align more closely with the granularity of tool call events.

Mastra and AI SDK lean more toward lightweight orchestration, with direct support for simple chaining and routing, while complex patterns require users to compose them themselves.

Mastra is an Agent development framework for TypeScript with a built-in evaluation module that is deeply coupled with the Agent runtime and can directly intercept and record detailed information about each tool call.