Sauna by Wordware

Positioning: an AI lab's flagship that is not open for business yet

I want to lead with the product state, because it determines everything else in this review. Wordware, the San Francisco AI lab founded in 2023 by Filip Kozera and Robert Chandler, has made Sauna its main product: an intelligent workspace built around what the company calls compounding context — an assistant that learns how you work, keeps persistent memory of your projects, and acts on busywork while you are away. The pitch is strong and the pedigree is real (Y Combinator, a $30 million seed round in 2024 led by Spark Capital with Felicis — which the company bills as the largest seed in YC history — and the #1 Product Hunt launch of all time for its earlier product). But as of my October 2026 check, Sauna is behind a waitlist: sauna.ai shows "Sauna is warming up—we'll call you in soon," with a Typeform for priority access. You cannot buy this product today.

What remains fully accessible is the earlier product, Wordware v1 — the natural-language agent builder that made the company's name — which is still live at app.wordware.ai as a self-serve product with published pricing. So this review covers both: the workspace most readers are here for (which I could only assess through its signup flow, published materials and the company's technical talks), and the legacy builder I actually used.

What Sauna is designed to do

Per Wordware's site and the company profile published by AI Engineer (checked August 2026), Sauna is a workspace rather than a chatbot. It connects to your email, documents and other tools, builds a persistent model of your work — stored, per the company, in a queryable file system rather than a flat chat log — and does knowledge work with it: drafting messages, preparing briefs, filing tickets, and capturing recurring tasks as reusable automations. Work is organized into "spaces" with their own rules, team memory is shared across members, and the human stays in the loop — you review, approve or edit what Sauna produced. It runs on web, iOS, Slack and other messaging channels, connects to more than 3,000 tools and apps, and accepts custom MCP connectors. The homepage framing is blunt: "Less busywork. More of whatever you'd rather be doing."

The design bet is that context, not model access, is the product. That is a defensible position — every serious AI assistant converges on the same frontier models, so the durable layer is what the assistant knows about you and how reliably it acts on it — and Kozera and Chandler have been making that argument publicly since their 2025 talks on documents as the programming substrate for agents. It is also a hard bet: persistent-context assistants live or die on trust, permissioning and error rates in exactly the areas (email, tickets, calendar) where mistakes are visible and costly.

What Wordware v1 actually does

Wordware v1 is the natural-language development environment the company started with: you describe an AI workflow in English, the platform builds it, and you deploy it as an API endpoint. The parts I appreciated when it launched still hold up: prompt version control (every change to the agent's instructions is tracked and revertible), a cloud IDE with a library of 2,000+ integrations and pre-built tools, and deployment that really is one click from the editor to a live endpoint. Pricing is published at wordware.ai/pricing, at time of writing: Builder at $0/month with $5 in monthly credits; Free at $0/month with 900 monthly credits and up to 3 deployed workflows; Professional at $39/month per seat with 4,500 credits and up to 10 deployed workflows; Business at $149/month with 18,000 credits, unlimited workflows and 5 included seats ($99 per additional seat); and Infrastructure at $1,999/month with 270,000 credits, 20 seats, custom tools and 24/7 support. SOC 2 and HIPAA compliance appear only at the top tiers.

My testing notes

On Wordware v1, two concrete observations. First, the natural-language build loop is real, not a demo trick: I described a text-classification-plus-notification workflow and had a runnable version with a deployed API endpoint in roughly fifteen minutes, with the prompt editable as versioned text rather than buried in configuration. Second, the credit model behaves predictably but punishes chatty agents: my test workflow burned credits per model call across its steps, and a multi-step agent exhausts the 4,500-credit Professional allowance much faster than the headline implies. Budget holders should model credits per run, not per user.

On Sauna, my testing is honestly bounded: I completed the signup flow (about 30 seconds, as advertised) and reached the waitlist state rather than the product. Everything I can report about its behavior — the context file system, spaces, automations, the 3,000+ integrations — comes from the company's own materials and its founders' conference talks, plus the AI Engineer profile. I have seen no independent review, no published benchmark, and no customer case study with numbers. When a product is waitlist-only, that is the correct state of evidence, and I am not going to pretend otherwise.

Limits — what this does not do

The first limit is access: Sauna cannot be evaluated, bought or compared today, and waitlist timing is unannounced. If you need an AI workspace now, this review cannot tell you Sauna beats the incumbents; nobody outside the company can. The second limit is track record: Wordware has pivoted once already (from the v1 builder to Sauna in 2025), and v1 users are effectively on a legacy product. The v1 builder, while still live, is described by the company's own navigation as "wordware v1" beside "Sauna OS" — the roadmap points away from it. Anyone building production agents on v1 today is building on a platform the vendor has moved past.

Third, the v1 product's ceiling: it is text-centric agent building — fine for document, classification and workflow tasks, but not a data platform, not an enterprise governance layer, and thin on evaluation tooling beyond version control. SOC 2 and HIPAA compliance are gated behind the $1,999/month tier. Fourth, for Sauna specifically the unanswered questions are the important ones: what permissions does it need over email and documents, what happens when it drafts something wrong and you miss it, and what does it cost? None of these are published.

One caveat on the legacy product: budget credits per workflow run rather than per seat, and confirm support terms before you commit — the compliance and success-management features live at the $1,999 tier.

Verdict

There are two verdicts here. For Sauna: the concept is one of the most interesting in the workspace category, the founders have a genuine point of view about context as the product, and the integration surface (3,000+ tools, MCP connectors, Slack/iOS) is ambitious — but the product is waitlist-only, so anyone can promise anything. Rated 2.9 on the rubric, weighted down by a core-functionality score I cannot verify hands-on (40% of the rubric rests on evidence I do not have), middling value while pricing is unpublished, and a reliability picture made of pedigree rather than proof. For Wordware v1: a capable, honestly priced agent builder that I can recommend today for prototyping and small production workflows — but it is the legacy line, and that fact caps the rating more than the software does.

Sources

• https://sauna.ai/
• https://www.wordware.ai/
• https://www.wordware.ai/pricing
• https://www.wordware.ai/story
• https://ai.engineer/orgs/wordware
• https://dealroom.co/companies/wordware
• https://everythingxiaomi.com/compare/smythos-vs-wordware/