A bigger brain, a fresher engine, and a quieter release - Elyra v0.9.5

Some releases are loud. This one is more like tidying the workshop and then realizing you also built something genuinely useful while you were in there. Elyra 0.9.5 is small on the surface and meaningful underneath. Let's walk through what changed, why we did it, and how to actually use it.

The headline: MiniMax M3, straight from the source

The marquee addition in 0.9.5 is MiniMax M3 on the two direct MiniMax providers — minimax (global) and minimax-cn (China). No gateway in the middle, no extra hop. Elyra talks to MiniMax's own Anthropic-compatible endpoint.

Why we're excited about it:

• 512k token context. That is enough to drop a substantial chunk of a codebase into a single session and still have room to think.

• 128k max output. Big refactors and long files don't get cut off mid-thought.

• Reasoning built in. M3 is a reasoning model, so Elyra's thinking levels map cleanly onto it.

• Vision. It takes text and image input, so screenshots and diagrams are fair game.

• Friendly pricing. Roughly $0.60 per million input tokens and $2.40 per million output. For a 512k-context reasoning model, that is easy to like.

How to use it

Set your key and start Elyra:

export MINIMAX_API_KEY=your-key-here
elyra

Then pick the model from inside the session:

/model

Search for MiniMax-M3, hit enter, and you're on it. In China? Use MINIMAX_CN_API_KEY instead and you'll get the minimax-cn provider pointed at the right endpoint automatically.

The nicest part is that you don't have to commit up front. Switch mid-session whenever the task changes:

> Read everything under src/billing and explain how invoices flow through the system
  ... (Elyra reads, traces, summarizes)
/model            # jump to MiniMax-M3 for the heavy lifting
> Now refactor the invoice pipeline to emit domain events, and keep the tests green

Because M3 carries 512k of context, that first "read everything" step doesn't quietly fall off the back of the conversation when you get to the refactor. It's still there.

And since M3 accepts images, this works too:

> Here's a screenshot of the broken layout. Match the spacing to the design.
[attach screenshot]

The fresh engine: Anthropic SDK 0.97.1 → 0.102.0

This one looks like plumbing, and it mostly is — but there's a nice connection worth pointing out.

MiniMax's direct API is Anthropic-compatible. Elyra speaks to minimax and minimax-cn through the same anthropic-messages path it uses for Claude. So keeping @anthropic-ai/sdk current isn't just about Claude models; it quietly keeps the MiniMax direct providers healthy too.

We bumped the SDK five minor versions, ran the full type-check and the Anthropic SSE-parsing tests, and confirmed nothing in the provider layer needed changing. From your side, this is invisible — which is exactly the point. The best dependency upgrades are the ones you never notice.

A little housekeeping (the unglamorous hero)

While preparing this release, we hit a snag: the publish pipeline aborted before it even started building. The culprit was our native clipboard package, which had quietly shipped without a clean script. Every npm run clean --workspaces tripped over it.

So 0.9.5 also includes a tiny, deeply unexciting fix: clean and check scripts for @elyracode/clipboard. It's the kind of one-line change that nobody writes a blog section about — except we just did, because it's a good reminder that smooth releases are built on boring correctness. Future release runs won't stumble on this again.

How to upgrade

If you've got Elyra installed globally:

npm install -g @elyracode/coding-agent

Or, the way we like best, from inside a running session:

/update

It updates Elyra and restarts without making you leave your conversation.

The short version

• New: MiniMax M3 on the direct minimax and minimax-cn providers — 512k context, 128k output, reasoning, and vision, at a comfortable price.

• Changed: @anthropic-ai/sdk upgraded to 0.102.0, keeping both Claude and the MiniMax direct path on a current foundation.

• Fixed: Release tooling no longer trips over the clipboard package.

Nothing here demands your attention, and that's a feature. Set a key, run /model, and let M3 chew on something big. As always, the CLI stays free, runs locally, and calls providers directly — no account, no proxy, no telemetry.

Happy building.