MMastra

Replit’s Agent 3 spins up thousands of Mastra agent sandboxes every week, Salesforce built a harness for 100,000 developers on it, and Marsh McLennan’s Mastra-based search tool is used by more than 100,000 people daily, according to Mastra. Those are big names for a framework whose company raised its $22 million Series A only in April 2026. Mastra is an open-source TypeScript framework for agents and workflows, free under Apache 2.0, with a hosted platform that starts at $0.

It is best for JavaScript and TypeScript teams that want to build agents into web products with memory, evals and tracing included. It is built for TypeScript, and the hosted platform bills on several separate meters. The framework and Starter plan are free, Teams is $250 a month, and we rate Mastra 9.0/10 on our scale. We have not used Mastra ourselves; this review is based on Mastra’s website, pricing page, funding announcement and Product Hunt reviews.

Customer Evidence and Price Data Mastra Publishes

Mastra publishes an unusual amount of customer evidence for a young company, alongside a detailed price list.

For a buyer, the customer list is the strongest signal. Replit and Salesforce are themselves agent builders, and running thousands of sandboxes a week or serving 100,000 developers is a production workload, not a pilot. These are vendor-published claims rather than independent audits, but they are specific and checkable with the companies.

The funding tells a second story. A $35 million total is small next to Google, OpenAI or Microsoft, which back competing frameworks. Mastra’s answer is to be the TypeScript-native option and to earn revenue from the hosted platform, which is why the platform has so many usage meters.

A TypeScript Library With Optional Hosted Services

Mastra is a TypeScript library you add to an application, plus optional hosted services for running, observing and storing agents.

Agents

• Model router: one interface for more than 1,000 models from providers including OpenAI, Anthropic and Google, so switching models is a configuration change.
• Tools and MCP: agents call tools, and tools can be shared across agents through MCP.
• Memory: message persistence plus observational memory, which learns about users automatically.
• Tool approval: a person or an external system can approve or reject a tool call before it runs.
• Supervisor agents: a lead agent coordinates specialist agents for multi-agent work.
• Workspaces: filesystem access, sandboxed code execution and reusable skill files.

Workflows

For work that must follow a set path, workflows run sequential steps, parallel branches, conditions and loops. A workflow can suspend for human input and resume later with its state persisted, and time-travel replay lets developers rerun from an earlier step to debug. Agents can sit inside workflow steps, so model-led and fixed logic mix in one flow.

Evals, Tracing and Studio

Mastra includes guardrails, scorers (model-graded, rule-based and statistical), tracing, and input and output processors that help defend against prompt injection. Mastra Studio is an interactive UI for running agents and workflows, swapping models and tracing execution; during development it runs locally. Non-technical colleagues can use it to edit prompts and create datasets, and it tracks token usage, latency, error rates and costs. Since April 2026 Studio is also available in the cloud or self-hosted, alongside Mastra Server for deployment and a Memory Gateway that other frameworks can use.

Starting a Project and Wiring In Model Keys

1. Create a project: add Mastra to a TypeScript project or start a new one.
2. Connect models: add provider keys, or use the hosted gateway.
3. Build agents and workflows: define instructions, tools, memory and workflow steps in code.
4. Test in Studio: run agents locally, inspect traces and build evaluation datasets.
5. Deploy: ship to any Node.js-compatible runtime, including Bun, Deno and Cloudflare, with adapters for Express or Hono, or use the hosted platform.

Developers Own Approvals, Scorers and Guardrails

Developers write the agents, tools and workflows, decide which tool calls need approval and maintain scorers and guardrails. Someone has to watch traces and costs in production. Business users can tune prompts and datasets in Studio but cannot build new agents without code. Teams that self-host also own upgrades, database backups and scaling, which the hosted platform otherwise handles for a fee.

Free Framework, Optional Hosted Tiers

The framework is free under Apache 2.0. The hosted platform is optional; prices below are from Mastra’s pricing page in October 2026.

Other meters on Starter and Teams include retrieval storage ($20 per GB after 250 MB or 1 GB), Postgres compute ($0.50 or $0.40 per hour after 5 hours), Postgres storage ($1 or $0.75 per GB-month after 1 GB) and LibSQL rows written ($2.50 or $2 per million after 1 million).

From a Side Project to Production on Mastra’s Rates

These scenarios use Mastra’s published prices. Usage levels are our own illustrative assumptions, and model tokens are extra unless noted.

• A side project on Starter: 300,000 observability events and 60 CPU hours a month. The extra 200,000 events cost $20 and the extra 36 CPU hours $12.60, for $32.60 a month.
• A product team on Teams: 3 million events, 400 CPU hours and one always-on server. The plan is $250, the extra 2 million events $160, the extra 150 CPU hours $37.50 and the persistent server $100, for $547.50 a month.
• Model spend through the gateway: an app that would spend $1,000 a month on tokens pays $1,055 through Mastra’s gateway. Bringing your own key avoids that markup.

Self-hosting the free framework avoids platform fees entirely; you then pay your cloud provider and model vendors directly. Variable costs on the hosted side include egress, retrieval storage, Postgres compute and memory tokens, each with its own allowance.

Mastra Compared With LangGraph, OpenAI Agents SDK and CrewAI

All four are free, open-source developer frameworks. Language and how much hosted tooling each sells are the main differences.

For a team whose product is written in TypeScript, Mastra removes the need for a separate Python service just to run agents, and its Studio and memory features are part of the package. LangGraph is the closest rival with a JavaScript version and deeper graph control. The OpenAI Agents SDK and CrewAI are Python-first. Teams on Google Cloud may also compare Google ADK, which has a TypeScript SDK. More options are in our agent frameworks and open-source agents guides.

TypeScript Teams Building Agents Into Web Apps

• TypeScript and JavaScript product teams: agents live in the same codebase and runtime as the web app, with adapters for Next.js, React, Express and Hono.
• Startups that want a free path to production: the framework and Starter plan cost nothing, and Teams is a flat $250 plus usage.
• Teams that need memory and evals built in: observational memory, scorers, guardrails and tracing come with the framework.
• Companies mixing engineers and domain experts: Studio lets non-technical colleagues edit prompts and build datasets.
• Enterprises with data residency rules: self-hosted Enterprise keeps traces, prompts and outputs inside your VPC for a flat annual fee.

Python Shops and No-Code Users

• Python-only teams: Mastra is TypeScript; LangGraph, CrewAI, the OpenAI Agents SDK or Google ADK fit better.
• Non-technical users: building agents takes code; a no-code agent builder is a better fit.
• Buyers who want one predictable bill: the hosted platform meters events, CPU, memory tokens, storage, databases and egress separately.
• Teams that need a large vendor behind the framework: Mastra is a venture-backed startup with $35 million raised, against cloud giants backing rival frameworks.
• Teams that only need one simple agent: a single model call with a tool or two may not justify adding a framework, a studio and a hosted platform to your stack; start smaller and adopt Mastra when workflows, memory and evals become real needs.

Events, CPU Hours and Other Meters to Track

• Several meters: track events and CPU hours in Studio during testing; overage on Starter is $10 per 100,000 events and $0.35 per CPU hour.
• Gateway markup: routing model calls through Mastra adds 5.5% to token costs; bring your own key if spend is large.
• Retention limits: Starter keeps data for 15 days, which may be too short for debugging or audits.
• Learning curve: one Product Hunt reviewer found Mastra hard for beginners and asked for simpler setup and more examples.
• Vendor-reported results: customer figures such as Marsh McLennan’s 100,000 daily users come from Mastra.
• Fast-moving product: the platform launched in April 2026, so features and prices may change; check the pricing page before committing.

A $22 Million Series A Led by Spark Capital

• April 9, 2026: Mastra raised a $22 million Series A led by Spark Capital, bringing total funding to $35 million, and launched Mastra Studio in the cloud or self-hosted, Mastra Server and a Memory Gateway.
• October 2026: Mastra’s pricing page listed a $0 Starter plan, a $250-a-month Teams plan and a flat-fee self-hosted Enterprise option.

Scored 7.9 for TypeScript Teams Shipping Web Products

For TypeScript teams building agents into web products, yes. Mastra scores 9.0/10 on our scale: strong on pricing value, integrations and autonomy, with positive but limited user reviews. The framework is free, includes memory, workflows, evals and a studio, and has real production users such as Replit, Salesforce and Marsh McLennan.

The points to plan for are the TypeScript focus and a hosted platform with several usage meters. Start with the free framework and Starter plan, run a real workload for a month and read the event and CPU totals in Studio. If the bill looks unpredictable, self-host or bring your own model key. Python teams should look at LangGraph or Google ADK instead.

Does Mastra Work With Python, and What Teams Costs

Is Mastra Free?

Yes. The framework is open source under Apache 2.0 and free to self-host. The hosted platform’s Starter plan is also $0 a month with 100,000 observability events and 24 CPU hours, with usage charges beyond that.

How Much Does Mastra Teams Cost?

Teams is $250 a month and includes 1 million observability events, 250 CPU hours, 6-month retention, SSO and SOC 2 documentation. Extra events cost $8 per 100,000 and extra CPU time $0.25 an hour.

Does Mastra Work With Python?

No. Mastra is a TypeScript framework that runs on Node.js-compatible runtimes such as Node.js, Bun, Deno and Cloudflare. Python teams usually choose LangGraph, CrewAI, the OpenAI Agents SDK or Google ADK.

Which Models Can Mastra Agents Use?

Mastra’s model router covers more than 1,000 models from providers including OpenAI, Anthropic and Google. You can use your own provider keys or Mastra’s gateway, which charges market token rates plus 5.5%.

Who Uses Mastra in Production?

Mastra names Replit, Salesforce, Marsh McLennan, Brex, Sanity, Factorial, Indeed, MongoDB and Workday among its users. It says Replit’s Agent 3 spins up thousands of Mastra sandboxes a week and Marsh McLennan’s Mastra-built search tool serves more than 100,000 people daily.