EN
Back to the archive

The archive · Developer & Business Tools · Product decision · 2023

Hegel AI's PromptTools bet: code-first LLM evaluation drew 211 HN points, 3.1k stars

Hegel AI built open-source prompttools for evaluating LLMs and vector DBs in code; its 2023-08-01 Show HN drew 211 points and the repo later passed 3.1k stars.

Hegel AI

The betThat evaluation could be code-first and self-hosted: short Python experiments across models, prompts and vector DBs run locally, then turn into unit tests and CI/CD.Live

What the business is

Open-source, self-hostable Python tools (prompttools) for experimenting with, testing and evaluating LLMs, vector databases and prompts — the product wedge Hegel AI shipped into LLM evaluation.

How it started

Kevin and Steve said in their Show HN that they had started a month earlier, in July 2023. Steve had worked at Google with HuggingFace and Lightning on running the newest open-source models on TPUs; he saw the community contributing powerful models that were still hard to discover and evaluate — it was not clear when Llama or Falcon could replace GPT-4. PromptTools was their attempt to simplify and scale that evaluation: write a short Python script that runs checks across models, parameters and prompts, then pass the results into an evaluation function for scores, all executed locally.

What happened

The 2023-08-01 Show HN drew 211 points and 24 comments. Engineers called it exactly the tooling one realizes is needed when starting to build apps on LLMs; commenters compared it with promptfoo and ChainForge and pressed on whether LLM auto-evaluation can be trusted. The founders answered with their roadmap — structured-output validation, semantic-similarity checks, manual feedback, multiple-LLM judging, exports to CSV, JSON, MongoDB and Markdown — and invited PRs for function calling and Qdrant support. The repo kept moving: 3.1k stars, 256 forks and 638 commits by the 2026-09-05 crawl.

No ending yet — it is still running.

Background

Kevin and Steve started PromptTools in July 2023, a month before their Show HN, under the Hegel AI name: open-source, self-hostable Python tools for experimenting with, testing and evaluating LLMs, vector databases and prompts. The README's core idea is evaluation through familiar interfaces — code, notebooks and a local playground — with integrations for OpenAI, Anthropic, Google, HuggingFace and LlamaCpp plus vector stores such as Chroma, Weaviate and Qdrant.

Steve had worked at Google with HuggingFace and Lightning on running open-source models on TPUs. He saw the community shipping powerful models that were hard to discover and evaluate: it was not clear when Llama or Falcon could replace GPT-4. PromptTools was the answer — a short Python script runs checks across models, parameters and prompts, and results stay on the developer's machine, with no server forwarding and no collection of keys or outputs.

The 2023-08-01 Show HN drew 211 points and 24 comments. Engineers called it needed tooling, compared it with promptfoo and ChainForge, and questioned how far LLM-as-judge can be trusted; the founders responded with details on structured-output validation, semantic similarity, manual feedback and multiple-LLM scoring. By the 2026-09-05 crawl the repository showed 3.1k stars, 256 forks and 638 commits.

What has to be true

  • Dated, specific traction: the 2023-08-01 Show HN (item 36958175) drew 211 points and 24 comments, and the hegelai/prompttools repo showed 3.1k stars, 256 forks and 638 commits on the 2026-09-05 crawl.
  • A concrete founding bet stated by the builders: Steve saw powerful open models at Google that were hard to discover and evaluate, and PromptTools was built to simplify and scale that process.
  • The wedge is a deliberate inversion: evaluation runs in local code and notebooks rather than a hosted platform, with no server forwarding and no collection of keys or inputs.
  • The launch thread shows real pull: engineers called it needed tooling and debated LLM-as-judge circularity against promptfoo and ChainForge — the problem the founders said they were solving.

What can be applied

A founder's own unsolved pain is a credible wedge: Steve could not tell when an open model could replace GPT-4, so PromptTools made 'evaluate in code' a five-line habit that spread via open source.

Aftermath

As of 2026-09-05 the hegelai/prompttools repository is live with 3.1k stars, 256 forks and 638 commits; the README still positions it as open-source, self-hostable evaluation tooling from Hegel AI and links documentation, a hosted Streamlit playground and a Discord. The material records no funding, revenue or corporate outcome for Hegel AI, so nothing beyond the project's continued development is claimed here.

Sources

spotted an error? The archive wants to know.

Your turn

You just read one. Describe what you are building, and see who is betting on the same thing.

Free account · 3 free questions · no card

Related cases