anfra retweeted
Day 5 of building an open-source agentic BI in public. Today is a quick rant about {context,semantic} layer.
There has been a lot of discussions about semantic layer and context layer recently. And one argument that I came across is this: if the AI models are getting so good, do you really need semantic layers, or just the context layer is good enough?
Honestly, I don't understand why there's such an argument.
To me, semantic layers and context layers are 2 complementary things, they represent both the quantitative and the qualitative side of the business.
While the context layer obviously contains the messy business context across different documents, the semantic layer (aka the the metrics layer) represent the formularized, quantitative aspects of the business. And you need both in order to function well.
Not needing a semantic layers is like saying I can already describe my proof in English, why should I need to use some sort of mathematical formula with it?
Well of course you can describe your proof in English, but English (or any other human language) is an imprecise language, there can and will be different interpretation of the same thing. While not just use a indisputable, computable formula for it.
This is why in anfra we ship both a context layer with a powerful semantic search functionalities, and a semantic layer with an advanced semantic query language. Both are independent components that interact well together.
Am I missing anything of the arguments?
p/s: If you want to implement an BI that has the best practices of agentic analytics, give anfra a shot.
anfra retweeted
Day 4 of building an open-source agentic BI engine in public.
Over the weekend my colleague @imanthonytd built a prototype of anfra's desktop app, because… why not?
Here’s a rough demo (live recording).
Claude/Codex integration is done using using ACP, originally built by the @zeddotdev team. This way, users get AI directly from their Claude/ChatGPT subscription. They don't even need to install Codex or Claude Code app.
What do you guys think? Can this be the agentic alternative of Tableau Desktop or PowerBI Desktop?
anfra retweeted
Mom come pick me up they’re arguing about semantic layer vs ontology in the group chat
anfra retweeted
web developers have had DevTools for 20 years.
data teams still debug with screenshots in Slack.
so we built DevTools for BI at @anfraAI .
meet Inspect: point at any number, see the SQL that ran, trace it to the table.
your agents can get it too, via the CLI.
check out: demo.anfra.ai
anfra retweeted
Day 2 of building open-source analytics engine in public. Let’s talk about low-cost, open-souce data stack.
Since building @anfraAI I have been trying this setup out, and it has soon become my new favorite open-source/almost $0 agentic data stack.
1. @duckdb (self hosted) or @motherduck free plans for data warehouse.
2. @dltHub and @getdbt for ELT (open-source). If it’s simple tables/sources, I might just skip them or hand-roll something with agents.
3. AMQL (amql.org) as a semantic layer, metrics definition. It ships with anfra. It is static-typed with composable/programmable consturcts (not YAML-based strings).
4. anfra as the open-source agentic analytics engine + custom BI apps frontend. Either self-host or free cloud.
The main cost is just usually hosting, which I typically go with a @Hetzner_Online VPS box ($7/mo), best bang for the bucks I can find.
If the data volume is low, I just go with Motherduck free plan (10 GB free storage) + anfra Cloud (free plan) to save the management hassle.
Contexts are either markdown files, dbt descriptions or Notion, that get ingested into anfra via anfra context layer.
All this is set up with coding agents (Claude Code, Codex or @opencode). I rarely look at the code anymore.
What you get:
- Interactive, customizable dashboards on top of a governed analytics backend
- Everything stored as code, checked in to Git
- Connect ChatGPT/Claude to anfra MCP and you get AI analytics with trusted metrics for free (no extra API token cost)
What does your low-cost agentic data stack look like?









