Mastra
An open-source TypeScript framework that developers use to build, run and observe AI agents and workflows, with a hosted Studio and cloud deployment option.
What it does
- Agent orchestration
Used for
- Workflow automation
Features 7
- AgentsTyped agents defined with instructions, models, tools and runtime behavior in one place.
- HarnessCoordinates multi-mode agents with shared state, threads and storage; the same harness that runs Mastra Code.
- MemoryDurable agent context with observational memory, semantic recall and thread-aware storage.
- SignalsSends input to an agent's thread at any time, including while the agent is working.
- Observational MemoryKeeps a dense observation log, maintained by background agents, that replaces raw message history as it grows.
- Code modeRuns multi-tool computations in an isolated sandbox and returns the result as a single response.
- EvalsScores agent outputs with automated tests that use model-graded, rule-based and statistical methods.
How it compares
-
CrewAI OSS
LangChain's own 'The best AI agent frameworks in 2026' page names Mastra (16 mentions), CrewAI (14) and LangGraph (16) as the frameworks compared.
Company disclosure · 20 Sep 2026 -
LangGraph
LangChain's own 'The best AI agent frameworks in 2026' page names Mastra (16 mentions), CrewAI (14) and LangGraph (16) as the frameworks compared.
Company disclosure · 20 Sep 2026
Sources
- Pricing
-
Framework cards: Free "$0/ month", "Apache 2.0 licensed"; Enterprise "Custom pricing", "Licensed self-hosted features" (RBAC, SSO, IAM, network policy, data in the customer's VPC) for "a flat annual fee with no per-trace or per-seat metering". The homepage: "Enterprise features are source-available under the Mastra Enterprise License."
- Description
- What it does
- Used for
-
Official documentation · 24 Oct 2026
'Orchestrate multi-step workflows with branching, snapshots, suspend and resume, and scheduled runs.' (https://mastra.ai/llms.txt)
- Sold within
-
The Free card's "Get started for free" links github.com/mastra-ai/mastra; the core is Apache 2.0 and installs from npm.
- Name
- URL
- Interfaces
-
Company disclosure · 8 Oct 2026
Quickstart is an npm/CLI scaffold (
npm create mastra@latest) plus a local Studio web UI opened bymastra dev.
Also from Mastra
-
Mastra Studio Developer tool
A shared workspace where a team inspects agent runs through traces and cost, latency and quality metrics, scores outputs with LLM-as-judge or code, runs dataset experiments and versions prompts and agent configurations, open to non-engineering reviewers without source access.
-
Mastra Platform Observability Infrastructure service
Hosted telemetry store for Mastra agents in production: it keeps the metrics, structured logs and step-by-step traces of model calls, tool calls and workflow steps for every run, with dashboards for cost, latency and quality scores.
-
Mastra Server Infrastructure service
Hosted deployment target that runs a Mastra application's agents, tools and workflows as an API server from one command, with routing and scaling handled by Mastra and billing by CPU time and data egress.
-
Mastra Gateway API service
An OpenAI- and Anthropic-compatible model gateway that routes one API key to 300+ models with provider failover, and adds observational memory that compresses each conversation into notes so agents keep context without rebuilding it.
-
Mastra Agent Builder Agent platform
A UI mounted inside a team's own Mastra server for creating and operating internal agents, with multi-tenant workflows, role-based access and channel integrations, licensed for production under Mastra's enterprise licence.
-
Mastra Factory AI agent
An open-source software delivery environment where coding agents take a connected repository's issues through investigation, planning, implementation and review to a reviewed pull request, with people deciding at each step; self-hosted or deployed on Mastra's platform.
-
Mastra Code AI agent
An open-source terminal coding agent with Build, Plan and Fast modes and explore, plan and execute subagents, whose observational memory compresses the session into notes instead of compacting it, so long sessions keep their context.
Something wrong here? Send a correction — quote this product id: mastra-framework.