We're excited to introduce Claude Code, Cursor, and Codex agents in Mastra. Starting today, Mastra is now a meta-harness:

Mastra has always been model-agnostic. Today with so many harnesses, we're taking a step towards being harness agnostic as well.

Over the last couple months, we noticed a number of our users delegating tasks from their Mastra agent to a lab harness. So we wanted to make this an official, first class supported behavior.

Harnesses maintain their existing feature set — file handling, codebase reasoning, multi-agent handoffs, permission controls.

They also inherit Mastra's agent surface: Calls: .generate(), .stream() etc Composition: workflow steps, subagent delegation, attached tools Studio Observability: logs, traces, token usage, cost Per-run options: requestContext, structuredOutput, abortSignal, instructions, etc

Full blog post:

Lmao I literally just moved my code over to vercel ai sdk to do this Guess I’m moving back to Mastra

the funny thing is we'd actually shipped it a week ago, just hadn't announced it yet

opencode when

soon

The harness layer is useful but the real leverage comes from treating agent sessions as first class observable infrastructure rather than one off conversations that disappear when the tab closes.

yup!

bring your own harness!

excellent

ty!

This is tuff

Not far enough though. Need codex auth support, Opencode Go auth support, ship your own harness with plugins/extensions e.g subagents, compaction. If mastra doesn't start accumulating plugin community, it will not be relevant player in harness ecosystem. For now switching to pi

yeah we are working on this

元马具这个说法有点上头

ty!

Does this mean, I can now use my claude code subscription, or is this still claude -p?

you may have about as much luck as we do getting answers from anthropic about this

will chatgpt pro subscriptions work with the codex harness?

they should!

Using the AI sdk for a project cause I needed better control of the full flow. (Ie: using Honcho from Hermes, or pragmatically injecting tools based on discovered skills). Why should I switch?

you can do everything you're already doing plus you also get memory, evals, datasets, studio, etc

pi/hermes on roadmap?

yup!

need hermes integration thanks

🫡

@zeeg Very nice!

@zeeg ty!

Long time coming ✊

yup

Omg

👏👏👏

🙌

Wait damn what happens to Mastra The last time I used it was like early 2025 or smth that’s crazy

life moves fast in the fast lane

Сool! Also Sam, inviting you to the beta test! I'm building which helps you grow on X and acquire users

Beautiful 😍

lol 😂

This is where a tiny desktop status layer becomes useful. Once Claude Code/Codex agents run inside a meta-harness, I still want running / waiting / stuck visible outside the app. I built Agent Island for that notch-level signal.

not sure 'meta-harness' clicks yet, cursor and claude code are both pretty opinionated about their own context. does running them inside mastra actually reduce friction or do you end up juggling two context windows?

generative ui inside an isolated sandbox via experimental harness agents is a massive leap forward. standard stream completion usually breaks the moment structural layout shifts, but pushing decoupled json-render schemas guarantees layout predictability at the runtime boundary. are you executing the component serialization deterministically, or relying on reactive token parsing to catch schema regressions before the ui layer mounts?

Pitch me Mastra over vercel ai sdk

has everything you need to get into production. memory, studio, observability, workflows, evals, datasets, etc

yessss

🙂

Excuse me what!

yup!

love mastra and respect the direction here. i’m still a bit skeptical though, would love to see the actual user cases where people want claude code/cursor/codex inside a mastra workflow

some people want to mix it up

harness-agnostic is the right call, nobody wants to bet their whole stack on one lab's cli. the part i'd watch is the handoff: when mastra delegates to claude code then picks it back up, does context survive, or does each harness start cold and re-discover the repo?

yup!

great job Sam

@DanielWhit21874 ty!!

🤍🤍

HUGE

Hey folks, thanks for the updates! About long-running workflows, is there any integration w/ @temporalio? <3

sorry dumb question - does this mean I can pass my claude / openai oauth key (or let my users do that if they trust me) and piggyback off their max/pro sub?

Mastra is the way 🗿

Cool! Even I'm building a mera harness for all models but self improving at @anymo_ai

@anymo_ai nice!