Mastra is the TypeScript-native agent framework built by the former Gatsby team. It gives JavaScript and TypeScript teams agents, graph-style workflows, RAG, memory, evals, observability, MCP support, and deployment paths without moving the agent layer into Python.
System Verdict
Pick Mastra if your product stack is TypeScript and you want agent logic to live beside the rest of your app. It is strongest for teams shipping AI features in Next.js, React, Node, Hono, Astro, serverless, or standalone API services.
Skip it if your organization is Python-first or wants a visual no-code builder. LangGraph, Microsoft Agent Framework, CrewAI, LangFlow, and Voiceflow are better fits depending on whether the gap is language, enterprise cloud alignment, or builder UX.
Best plan: start with the mostly Apache 2.0 framework and $0 Platform Starter. Move to Teams at $250/month only when shared Studio/Server deployments, SSO/SOC 2 docs, longer retention, or higher included CPU/egress/storage matter. Budget separately for CPU, observability, egress, persistent server uptime, Memory Gateway, model-provider costs, and any sandbox or ACP harness infrastructure.
Key Facts
| Last verified | June 22, 2026 |
| Language | TypeScript / JavaScript |
| Framework license | Core framework and most repository code under Apache-2.0; ee/ code uses Mastra Enterprise License |
| Company | Mastra, founded in 2024 by the former Gatsby team |
| Current site signal | 25.3k GitHub stars displayed on mastra.ai/GitHub |
| Funding/team signal | About page says $35m raised and a globally distributed team in the mid-30s |
| Core primitives | Agents, tools, workflows, RAG, memory, evals, observability, MCP, Browser, Server, Studio, model routing |
| Platform Starter | $0/month with unlimited users, deployments, and projects plus included usage allowances |
| Platform Teams | $250/month with multiple teams, SSO, SOC 2 docs, 250 CPU hours, 100GB data egress, 100GB platform storage, and six-month retention |
| Enterprise | Custom platform pricing and custom licensed self-hosted features |
What Changed Recently
- June 22, 2026 pricing recheck: Starter remains $0/month, Teams remains $250/month, Enterprise remains custom, and self-hosted Enterprise remains custom. The live page still makes usage math central: observability events, CPU, egress, persistent servers, Memory Gateway input/output at market rate plus 5.5%, memory tokens, retrieval storage, platform storage, LibSQL rows, Postgres compute, database storage, and model-provider costs can all move the bill.
- Mastra’s traction claims moved. The site and GitHub now show roughly 25.3k stars, while the About page says Mastra raised $35m, gets millions of downloads each month, and has a globally distributed team in the mid-30s. Treat these as current marketing/procurement signals, not independent adoption proof.
- Code Mode is the new analytical-agent feature to test. Mastra’s June 9 post says Code Mode lets agents run multi-tool computations in an isolated sandbox, reducing repeated tool-result context and improving analytical answers. That makes sandbox policy and data-access boundaries part of the buyer review.
- Agent Signals and ACP widen the orchestration story. The June 3 Signals post adds steering, observing, notification, and state primitives for long-running agents; the June 2 ACP post adds delegation to Claude Agent, Codex CLI, Cursor, Gemini CLI, or compatible harnesses. Useful, but permissions and workspace isolation need a real security review.
- The model router is bigger than the old page implied. Current docs claim 4,539 models from 133 providers through one interface, with fallbacks and provider-specific options. That is a flexibility advantage, not bundled model spend.
- Mastra Platform is now the commercial center. Studio and Server can be deployed to the cloud, while the open-source framework remains free.
- Pricing is no longer just “$0 or $250.” The public page now exposes add-on rates for observability events, CPU time, data egress, persistent server uptime, Memory Gateway tokens, and retrieval storage.
- Memory Gateway matters beyond Mastra apps. Mastra markets it as memory for any agent/framework, with token and storage meters.
- Recent production updates focused on operations. The May blog cadence emphasized background tasks and observability work, which fits Mastra’s buyer: teams shipping agents as production services rather than one-off prompts.
- Model choice is intentionally external. Mastra is the framework and platform layer. Budget model-provider spend separately and avoid assuming a specific Claude, OpenAI, or Gemini model is bundled into Mastra pricing.
When to Pick Mastra
- TypeScript-first teams. Tool definitions, workflow state, agent inputs, eval outputs, and app code can stay in one language and type system.
- Agent APIs inside web products. Mastra can run inside existing frameworks or as a standalone server, so the agent layer can ship as normal infrastructure.
- Workflows need explicit control. Mastra workflows support graph-style control flow for multi-step processes where you do not want a fully autonomous loop.
- Production visibility matters early. Studio, evals, traces, logs, metrics, and CloudExporter make it easier to debug before the agent becomes a black box.
- Memory is a first-class requirement. Mastra’s docs include message history, resource/thread scoping, and observational memory for long-running conversations.
When to Pick Something Else
- Python stack: LangGraph, Microsoft Agent Framework, or CrewAI.
- No-code visual agent builder: LangFlow or Voiceflow.
- Pure observability layer: Langfuse or Helicone can complement Mastra without replacing the framework.
- Azure-first enterprise: Microsoft Agent Framework is usually easier to justify if Azure AI Foundry is already the standard.
- Simple one-off LLM calls: direct provider SDKs or the Vercel AI SDK may be lighter.
Pricing
| Plan / meter | Current public price | Notes |
|---|---|---|
| Open-source framework | $0 | Apache 2.0; build and host agents anywhere |
| Platform Starter | $0/month | Unlimited users/deployments/projects; includes 100K observability events, 24 CPU hours, 10GB egress, 15-day retention, and Memory Gateway starter allowances |
| Platform Teams | $250/month | Adds multiple teams, SSO, SOC 2 docs, 250 CPU hours, 100GB egress, 100GB platform storage, six-month retention, and larger Memory Gateway allowances |
| Platform Enterprise | Custom | Custom volume/retention, RBAC, audit logs, support and uptime SLAs, dedicated support engineer |
| Persistent Server | $100/project on Starter and Teams | Listed as 24/7 uptime add-on on the public pricing page |
| Usage add-ons | Varies | Observability events, CPU time, data egress, Memory Gateway input/output at market rate plus 5.5%, memory token usage, retrieval storage, platform storage, LibSQL rows, Postgres compute, database storage, and model-provider usage can add cost |
| Self-hosted Enterprise | Custom annual fee | Licensed self-hosted features, RBAC, SSO, IAM/network-policy integration, predictable annual fee, and data staying in the customer’s VPC |
Pricing verified June 22, 2026 against Mastra’s public pricing page. Model-provider costs (Anthropic, OpenAI, Google, and others) are not included in Mastra pricing.
Failure Modes
- Usage math is easy to underestimate. Teams that leave persistent servers running, generate high trace volume, or use Memory Gateway heavily may pay more than the headline $250/month Teams plan.
- TypeScript-only by design. This is a strength for web teams and a mismatch for Python-first data/agent teams.
- You still need production engineering. Mastra supplies framework and platform primitives, but auth, secrets, data retention, eval quality, and deployment architecture remain your responsibility.
- Sandbox and harness permissions are now part of the review. Code Mode, ACP delegation, and long-running Signals workflows are powerful, but buyers should check workspace isolation, tool approvals, external agent credentials, logs, and data boundaries before letting agents operate on production repositories or customer data.
- Fast-moving surface area. Platform, Memory Gateway, observability, background tasks, and integrations are evolving quickly; procurement teams should re-check docs before committing.
- No-code buyers will bounce. Studio helps inspect and test, but Mastra is not a drag-and-drop business automation builder.
Against the Alternatives
| Mastra | LangGraph | Microsoft Agent Framework | CrewAI | |
|---|---|---|---|---|
| Primary fit | TypeScript agent apps | Stateful production agents | Azure/Microsoft enterprise agents | Python role-based crews |
| Main language | TypeScript / JavaScript | Python + JavaScript/TypeScript | .NET + Python | Python |
| License posture | Apache 2.0 framework | MIT library | Open source | Open source |
| Hosted path | Mastra Platform: Observability + Studio + Server | LangSmith Deployment | Azure AI Foundry | CrewAI Enterprise |
| Best plan guidance | Start free, pay Teams when platform collaboration/usage limits matter | Free library, pay LangSmith when tracing/deployment matters | Use when Azure governance matters | Use for faster multi-agent crew prototypes |
| Watch-out | Usage meters beyond Teams | LangSmith deployment/trace costs | Azure ecosystem pull | Less explicit graph control |
Methodology
Produced by the aipedia.wiki editorial pipeline. Last verified June 22, 2026 against Mastra official site, documentation, pricing, model-router docs, platform/deployment docs, customers page, About page, GitHub repository, and recent Mastra Code Mode, Agent Signals, ACP, and harness blog surfaces.
FAQ
Is Mastra really free? The framework is free and Apache 2.0 licensed. Mastra Platform also has a $0 Starter plan, but production usage can create add-on costs once CPU, observability, egress, persistent server uptime, Memory Gateway, storage, or model usage grows.
Is all of Mastra Apache 2.0?
Not quite. The public repository says the core framework and most code are Apache-2.0, while directories named ee/ are source-available under the Mastra Enterprise License and require a valid enterprise license for production use.
What is Mastra Platform? Mastra Platform is the hosted product around Studio and Server. Studio is the hosted visual environment for testing agents, workflows, and traces; Server deploys the Mastra application as an API server.
How is Mastra different from LangGraph? Mastra is built for TypeScript product teams and web deployment ergonomics. LangGraph has the larger LangChain/LangSmith ecosystem and remains the stronger default for Python-first stateful agents.
Can I deploy Mastra without Mastra Platform? Yes. Mastra can run in Node-compatible environments, Bun, Deno, Cloudflare, standalone servers, web frameworks, VMs, containers, PaaS platforms, and cloud providers. The paid platform is a managed route, not the only deployment path.
What production companies use Mastra? Mastra’s customers page lists case studies and customer references including Sanity, Replit, SoftBank, WorkOS, Counsel Health, Factorial, Index, Cedar, and Kestral.
What changed for Mastra in June 2026? Mastra added ACP support on June 2, Agent Signals on June 3, and Code Mode on June 9. Those updates make it more relevant for long-running agents, coding-agent delegation, and multi-tool analytical workflows, but they also make sandbox, workspace, and permission controls more important.
Sources
Review History
- 2026-06-22: Re-verified platform pricing, usage meters, model-router scope, GitHub star signal, and June agent/harness posts. Updated model-provider count to 4,539 across 133 providers and added database meters to buyer cost warnings.
- 2026-06-15: Re-verified platform pricing, usage meters, model-router scope, deployment/platform docs, company traction claims, GitHub stars/license posture, customers page, Code Mode, Agent Signals, and ACP support. Updated watch-outs for sandbox execution, ACP harness delegation, provider-model costs, and enterprise-license boundaries.
- 2026-06-12: Re-verified Starter, Teams, Enterprise, persistent server, observability event, CPU, data egress, Memory Gateway token, Memory Gateway storage, platform storage, and retention meters. Corrected Starter observability allowance to 100K events and removed unsupported bundled-model implications.
- 2026-05-13: Re-verified Starter, Teams, Enterprise, persistent server, CPU, observability event, data egress, Memory Gateway storage, and retention meters against the live pricing page; no material changes. Added model-cost context for buyers using Mastra with hosted frontier-model APIs.
- 2026-05-10: Refreshed Mastra pricing, source provenance, platform/memory guidance, recent product changes, and buyer watch-outs.
- 2026-04-18: Initial TypeScript agent-framework review.
