The problem we kept running into

Most teams we talk to adopted AI the same way. Someone pays for ChatGPT, someone else for Claude, engineering uses a code assistant, and every one of those tools is a separate island. Agents read files, call tools and retry, so per-token bills are hard to predict. And outside engineering (sales, operations, support), AI barely shows up in daily work.

So we set three constraints for ourselves:

1. A team should share one workspace with AI, not ten personal chats.
2. The cost of a task should match the task. Summarizing an email doesn't need a frontier model.
3. If a provider goes down, rate-limits or changes its prices, the agents should keep working.

What Napsix is, concretely

Napsix gives you XIA, an assistant you talk to; Rooms, shared channels where people, agents and apps sit together; and agents you can delegate recurring work to, with permissions and schedules. Around that sit native apps: inbox (WhatsApp, Instagram, email and more), CRM, Docs, Sheets, Dashboards, Forms, Calendar, Storage, Automations and a site builder. It connects to 46 native integrations and 1000+ apps through connectors, including Mercado Libre, Mercado Pago, Tiendanube (where we're an official App Store app), Shopify, Gmail, Slack and Notion.

Chat commands do most of the work: /doc, /sheet, /task, /deal, /web, /summarize, /chart and others. They turn a conversation into an artifact or an action instead of a wall of text.

Bet #1: route by task, not by brand

Every request is routed to a model tier that fits it:

• Free / fast tier: open models such as Llama, DeepSeek and GPT-OSS.
• Premium: Sonnet, GPT-5-mini, Gemini Flash.
• Frontier: Claude Opus, GPT-5, Gemini 2.5 Pro, saved for work that actually needs them.

The tiers also do a second job: they are our availability story. There's always a free floor running, so if one vendor fails, the agent falls back instead of stopping. In practice, "which model is best?" matters less than "which model is enough for this step, and what happens when it's unavailable?"

Bet #2: own a model for the boring 90%

We trained and released Mato v1.0, our first model of our own. It's a fine-tune (SFT/LoRA) of a model from Alibaba's Qwen family, released under Apache-2.0, with 256k context and tool-calling tuned on Napsix's own tools. It runs on H100s through Daytona's Startup Grid Program, and we don't charge per token for it on any plan. It's on Hugging Face as german-gimenez/mato-v1.0-sft (gated, ES/PT/EN) and in our API as napsix/mato-v1.0.

Why bother? Office work is mostly repetitive, tool-heavy steps: draft a reply, file a CRM record, summarize a thread. A smaller model that knows your tool schemas well is often good enough, and it turns the free tier from a cost into something we control. The open question for us is evaluation: knowing when Mato is "enough" and when to escalate. That's where I'd love input.

Bet #3: humans approve, agents act

Agents in Napsix act in real systems, so permissions are per person and per agent. Sensitive actions can require human approval, and conversations and runs keep a history you can audit. Drafts are the default: XIA prepares the reply, the proposal or the document, and someone on the team confirms it. We think this is how AI gets adopted outside engineering. It doesn't replace anyone. It does the tedious first pass.

The stack

It's all built in-house on Next.js, Postgres and AWS Bedrock, with the Vercel AI Gateway behind a multi-model layer and our own integration platform. For developers we expose an OpenAI-compatible AI Gateway, a Crawl API (search/scrape) and an MCP connector, so Claude, ChatGPT or Claude Code can operate on Napsix data.

Some lessons so far:

• Tool schemas are the product. Good routing doesn't help much if the tools are ambiguous. Most of our quality gains came from tightening tool descriptions and outputs.
• Fallback must be boring. A fallback that changes the answer's format breaks downstream automations. We normalize outputs across tiers.
• Shared context beats a smarter model. A Room where the agent sees the thread, the docs and the CRM record usually beats a better model with no context.

Pricing, briefly

Napsix is freemium. The XIA Free plan costs USD 0 and needs no credit card: one user, 500 messages a month, one integration, and fast models without a cap. Paid plans start at USD 19/month, and there's a 14-day trial. Details are at napsix.com/pricing.

What I'd like feedback on

• How do you decide when a small, tuned model is "good enough" for agent steps? Evals, shadow traffic, user corrections?
• For those running multi-provider setups: how do you keep output formats stable across fallbacks?
• What would make you trust an agent to act in your CRM or inbox?

You can try it free at https://www.napsix.com/register. If you build with LLM agents, I'd genuinely like to hear how you handle these trade-offs. I'm happy to answer questions in the comments.

Germán Giménez, CTO & co-founder, Napsix (GOS SA), Mendoza, Argentina