# LearnHarness: a YouTuber with 30k subscribers used one video to get his open-source tutor 5500 stars · GitHub open-source project

> Source: https://stringers.io/en/github/gh_amosblomqvist__learnharness.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> Reach from the author's own YouTube videos: the previous video on the same topic got 1M views and sent about 4000 stars to his previous repo; this time it's the formal version for the same audience, and the name caught 'agent harness', which had been hot since September.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

amosblomqvist/learnharness · amosblomqvist

# LearnHarness: a YouTuber with 30k subscribers used one video to get his open-source tutor 5500 stars

**What it is** A free, open-source desktop AI tutor: it first tests where your understanding breaks, then walks you through a dependency graph step by step, runs practice in a separate session, and saves what you've learned as a local knowledge graph. Works with your own key, Claude Code, or an existing subscription.

**Why it took off** The audience from the author's own YouTube videos: his previous video on the same topic got 1M views and already brought his previous repo about 4000 stars; this is the official version for the same viewers. The name also hit "agent harness," which has been hot since September.

**C** **How you could follow** Don't build another general AI tutor. If you have an audience, copy the approach: validate an AI workflow with a video first, then open-source the companion tool and ship an installable release. If you don't have an audience, your only bets are the mobile sync, Obsidian integration, or Chinese version it hasn't done.

C grade: the stars came mainly from the author's own YouTube audience, which followers can't reach. TYSA, with the same idea, has only 21 stars, and the category still has DeepTutor at 55k stars and big-tech learning modes above it. The approach is worth learning; copying the project directly isn't worth it. What's going around the small circles

Data (updated daily) · Repo created 2026-10-16 · 6,138 stars · 7 days +1,633

GitHub · stars

Data 2026-10-24 → 2026-11-08 (16 days, 16 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

In 2026 Oct 23, Finnish YouTuber Eero Alvar posted a video titled "Agent Harness for Learning Things." The first line of its description reads "This is my full system for learning things with AI," followed by a GitHub link. The repo gained 278 stars the next day and 612 on the third day, reaching 5524 stars 21 days after creation. Copycat repos built on the same idea got only 21 stars in the same period. This project got its stars not from GitHub but from its author's own audience. This piece lays out the playbook and which step you can learn from.

### One video, one link

The video's chapters are the product's workflow: Goal, Understanding, Optimization prerequisite, Teaching, Practice, Model of your understanding. Judging from the chapter names, he demonstrates with a math goal related to optimization (inference): the AI doesn't lecture first; it asks layer by layer to find which prerequisite your understanding breaks on, then builds upward from that point.

As of Nov 6, the video had 467k views and 11k likes (YouTube, 11-06), while his channel had only 30.7k subscribers. Views at 15× his subscriber count mean the video was recommended widely, not just watched by longtime fans.

### What it is

A free, open-source (AGPL-3.0) desktop app with installers for Mac, Windows, and Linux, using AI as a tutor. The first screen of the README breaks it into four steps:

- **Measure** : Before a session, find out how far you understand each prerequisite. You answer questions, rate how confident you are, and use a "thinking box" to explain your reasoning in your own words.

- **Plan** : A dependency graph from where you are to your goal, visible to you, and it updates as you learn.

- **Teach** : Push one reasoning step at a time: guess first, then explain, then check. Fix misconceptions before building on them. The README's own words: the goal is for conclusions to feel "obvious, not just familiar."

- **Practice** : A separate practice session where questions don't hint at the topic, forcing you to recall which knowledge applies.

What you've learned is stored as a local knowledge graph of "propositions + inference chains," and the next session picks up from it. Model choice is yours: use your own API key, the Claude Code CLI you already have installed, an existing subscription, or a local model. No subscription fee, no markup. Chat logs and the graph are plain local files, and the system prompts can be read and edited.

### Why this one, and why now

The "AI tutor" idea is not new. HKU lab's DeepTutor has 55k stars; the Mr. Ranedeer prompt from 2023 years ago has 29.5k; ChatGPT, Claude, and Gemini all ship built-in study modes. LearnHarness shipped in three weeks because of three things followers find hard to copy.

**The audience was built beforehand.** About 8-14 months earlier, he posted "How I Use AI to Learn Things," which got 1M views. That video came with a repo called `learn`, described as "My AI learning system from this video," sharing his own setup as is. It went from 673 stars at 8-29 to 2970 stars at 10-01, and is now at about 4000. In other words, he had already tested the audience of "how to actually learn hard things with AI" and had already turned them into stars. LearnHarness is the official version for the same people: the previous one required building your own setup inside a coding agent like pi; this one works as a download.

**The name hit a buzzword.** In 2026 September, "agent harness" suddenly became a buzzword on X and in YC circles: YC ran a harness segment on Sep 7, and Greg Isenberg posted "AGENT HARNESSES ARE THE NEW GPT WRAPPERS" on Sep 13. Naming a tutor app "Agent Harness for Learning" makes developers understand it instantly and makes them more willing to click. How many stars the term added can't be isolated (inference: it acted as an amplifier only).

**For developers it's free.** It connects directly to the Claude Code CLI and subscriptions people already pay for. One blogger titled his post "Get a Free Claude Tutor" (chen.media, 11-01). Anyone already paying for an AI subscription can install it without spending another cent.

### Where the stars came from

| Date | Same-day star gains (Trendshift) | What happened |
|---|---|---|
| 10-16 | — | Repo created |
| 10-23 | 0 | YouTube video published |
| 10-24 | +278 | Entered Trendshift TypeScript daily chart at #15 |
| 10-25 | +612 | #17 on the site-wide daily chart and #3 on TypeScript; explainx posted "9 days, 1.5k stars," and reveneau followed up |
| 10-26 | +722 | Posted on Hacker News: 14 points, 0 comments |
| 10-27 | +712 | Korean blog and Spanish LinkedIn reposts; Chinese X aggregator accounts like GitHubDaily also posted (exact timing and engagement not verified) |
| From ¥10-28 | +535, then fell back to 180–240 per day | Long tail continues: 11-06 gained +181 that day |

The biggest peak tracks the video release almost exactly. The aggregator sites and repost accounts that followed amplified it; they weren't the starting point. Hacker News didn't help, and GitHub's official trending page didn't pick it up either (it doesn't list repos under about two months old). Everything came from off-platform.

There's another signal: 5524 stars with 866 forks, a fork-to-star ratio of 15.7%, above the roughly 10% common for new repos in the same period. With 232 open issues and 142 open PRs on a three-week-old desktop app maintained by one person, people are clearly installing and using it. The most-discussed issues include "Mobile app + sync," "Obsidian Support," and "Ability for Long term retention," along with "Has someone tested this on non math subjects" (GitHub issues, 11-06). Actual download numbers weren't found.

### What the copycats got

Someone has already copied it. Akas333/TYSA (Teach Yourself Anything) was created around 10-26 with a description almost word for word: "a learning harness that finds where your understanding…" It's also Electron plus a knowledge graph, and it reached 21 stars by 11-06. Other projects from the same period include dsh-aha-tutor, which wraps DeepSeek as a private tutor (27 stars), and socratic-workshop, a Socratic learning skill for Claude Code / Codex (23 stars).

The idea, the tech stack, and the name can all be copied. What can't be copied is the video with 460k views. **Most of this project's stars came from its author's audience, not from random GitHub users.**

### If it were you

**C.** Building "another LearnHarness" won't get stars; TYSA's 21 stars already show that. The category also has DeepTutor at 55k stars and big-tech study modes above it. What's worth learning is the playbook: validate demand with content first, open-source your own workflow, then ship an installable release with a second video. But this path requires that you already have an audience.

**Who it suits** : developers who already have tens of thousands of followers on YouTube, Bilibili, or X and teach "how to do X with AI"; or people willing to spend months on content before releasing a tool.

**Who it doesn't suit** : people without their own distribution channel who hope to earn stars through organic GitHub discovery; people who want to build another general-purpose version of the "AI tutor" idea; people counting on Hacker News to bring traffic (it got 14 points here).

**How to do it** : Pick an AI workflow you actually use and can explain (learning, writing, research, note-taking all work). First make a video about the method, and publish a repo called something like "the setup from my video" to see whether stars and comments show demand. If people want it, turn it into a desktop app anyone can download: installers for three platforms, auto-update, and the ability to use the user's existing subscription. Cover it on the README's first screen with the four-step flow and two images. Attach a term that's hot in developer circles to the name. Without an audience, the openings to try are the parts it hasn't done well: mobile with sync (top of the issues list), Obsidian integration, long-term memory review, and subjects beyond math; or a Chinese version that uses domestic models and is posted on Bilibili and Xiaohongshu. These gaps are inferences; nobody has verified they'd get stars.

**How to get users** : the main channel is your own video, with the repo link in the first line of the description; then let aggregator sites and multilingual repost accounts amplify it organically. In the Chinese-speaking community you can pitch repost accounts like GitHubDaily directly.

**Deadline** : "agent harness" is a hot term right now, with maybe a few months left. Once the term fades, the click boost from the name goes with it, but the "content-plus-tool" playbook doesn't depend on it.

**How this could be wrong** : (1) Someone without a built-in audience could build a learning agent app using only GitHub and X and hit 1000 stars within a month, which would show the category has natural demand and C underrated it. (2) Conversely, if the author's next repo with a video fails to take off, the playbook was a one-off. (3) If big-tech study modes add cross-session knowledge graphs and independent practice sessions, LearnHarness's differentiation disappears and there's even less room for followers.

### Data and sources

- GitHub: README and issues for amosblomqvist/learnharness (11-06); day-by-day stars and forks (recorded daily from 10-24, through 11-06); star counts for the author's other repos on his profile.

- Trendshift (trendshift.io/repositories/319002): day-by-day star gains and chart rankings from 10-21 to 11-06.

- YouTube: "Agent Harness for Learning Things" (published 10-23, 11-06 views at 467,566) and "How I Use AI to Learn Things" (about 8-14, about 1M views); channel @EeroAlvar.

- `learn` repo early star count: Wayback Machine 8-29 snapshot, and the 10-01 record.

- Secondary spread: explainx (10-25), reveneau (10-25), teckdeck (10-26), a Korean blog and a Spanish LinkedIn post (10-27), chen.media (11-01); Hacker News post at 10-26.

- "agent harness" buzz: YC 9-07, Greg Isenberg 9-13, omarsar0 X posts from mid-September.

- Competitor and copycat star counts on GitHub (around 11-06).

[GitHub repo ↗](https://github.com/amosblomqvist/learnharness)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| HKUDS/DeepTutor | Open source | 2025-12 | 55.4k stars, +5,749 in the last 7 days | A lab-built RAG + multi-agent study tutor. LearnHarness focuses on diagnosing where your understanding breaks and teaching step by step, and it's a desktop app built by one person. |
| Mr. Ranedeer AI Tutor | Open source | 2023-03 | 29.5k stars, stalled | Just a prompt, with no state, knowledge graph, or practice scheduling. |
| amosblomqvist/learn | Open source | 2026-08 | About 4,046 stars | The author's own earlier version, shared as a personal config as-is; it needs an agent environment like pi to run. LearnHarness is an app for ordinary users to install. |
| hhyqhh/inno-agent | Open source | 2026 | 1,381 stars | A learning companion built on the Pi SDK, featuring three-layer memory and IM channels, not diagnostic step-by-step teaching. |
| ChatGPT Study Mode / Claude Learning mode / Gemini Guided Learning | Big-company features | 2025 | Massive user base | Socratic chat; no explicit cross-session knowledge graph or separate practice sessions. Data lives in the cloud, and the prompts can't be changed. |
| Google Learn Your Way / NotebookLM | Big-company features | 2025 | Big | Rewrites your materials into textbooks or podcasts. It leans toward reading and doesn't diagnose understanding. |

Where users come from

- **Verified** **the author's own YouTube channel** : "Agent Harness for Learning Things" published 2026-10-23, 467k views and 11k likes, with the repo link in the first line of the description. Stars started climbing from 10-24 (Trendshift +278).

- **Verified** **audience built earlier** : "How I Use AI to Learn Things" reached 1M views, and its companion repo learn went from 673 stars (08-29) to 2,970 (10-01). Across his repos the author has about 11k stars in total.

- **Verified** **agent harness buzz** : 2026-09 YC harness special, Greg Isenberg's "AGENT HARNESSES ARE THE NEW GPT WRAPPERS", and back-to-back omarsar0 posts (x-radar).

- **Verified** **AI aggregator sites / SEO and multilingual reposts** : explainx 10-25, reveneau 10-25, teckdeck 10-26, Korean blog 10-27, Spanish LinkedIn 10-27, aifrontierpost 10-28, chen.media 11-01.

- Inferred **Chinese X repost accounts** : search snippets show GitHubDaily and Geek Lite posted about it on X. The timing and engagement of the original posts are unverified.

- **Verified** **Trendshift charts** : 10-24 to 10-29 it reached the TypeScript top25 (peak #3) and stayed in the site-wide top25 for 3 days.

- **Verified** **Hacker News** : one post on 10-26 got 14 points and 0 comments. It didn't take off.

## Why LearnHarness (amosblomqvist/learnharness) took off

> Channel: GitHub | repo 2026-10-16 | as of 2026-11-06:5,524 stars / 866 forks / 232 open issues / 142 open PRs (our GitHub repo_profile)

### Bottom line

This isn't a repo that took off through organic GitHub discovery. It's **a YouTube creator with 30k subscribers who open-sourced the companion tool for his second hit video** . Eero Alvar (GitHub amosblomqvist, Finland) posted "How I Use AI to Learn Things" two months earlier, which reached 1M views and brought about 4,000 stars to his previous repo `learn`. This time he turned his personal study workflow into a proper desktop app, released with the video "Agent Harness for Learning Things" (467k views in 13 days, 10-23), and stars began rising sharply the next day. The name also hit the "agent harness" term that has been heating up on X since September 2026.

### 1. What it is

An open-source (AGPL-3.0) Electron desktop app (macOS / Windows / Linux) that uses an AI agent as your tutor. The flow has four steps: **Measure** (probes how far your understanding goes, collecting answers, confidence, and self-reports from a "thinking box") → **Plan** (maps a dependency graph from your current level to your goal) → **Teach** (covers one reasoning step at a time: you predict, then explain, then check) → **Practice** (a separate practice session where questions don't name the concept being tested). What you've learned is stored as a local knowledge graph of propositions and reasoning chains, and you pick up where you left off next time.

- Bring your own key (30 and major model providers), **or connect the Claude Code CLI you already have installed, or a supported subscription** . Local models work too. No subscription and no markup. Data stays local as plain files. The system prompts are visible and editable.

- Source: README [github.com/amosblomqvist/learnharness](https://github.com/amosblomqvist/learnharness); website [learnharness.ai/](https://learnharness.ai/)

### 2. Growth data

| Date | Stars (our repo) | Trendshift daily star gain | Event |
|---|---|---|---|
| 10-16 | 0 | — | Repo created |
| 10-23 | — | +0(fork +14) | YouTube video "Agent Harness for Learning Things" published |
| 10-24 | 28 | +278 | Trendshift TypeScript chart #15 |
| 10-25 | 787 | +612 | Trendshift site-wide #17, TS #3; aggregator sites like explainx and reveneau start writing about it |
| 10-26 | 1,412 | +722 | Posted on HN: 14 points, 0 comments |
| 10-27 | 2,144 | +712 | Secondary spread on Korean and Spanish LinkedIn and similar channels |
| 10-28 | 3,269 | +535 | Our repo logged +1,125 that day. This likely reflects a different snapshot timing (Trendshift showed +535 that day), not an additional spike. |
| 10-29 → 11-06 | 3,647 → 5,524 | +390 → roughly 180–240 per day | Slowly tapering off, with a long tail still going |

- Our repo: gain_7d = 1,605, max_day_gain 1,125 (10-28, see the methodology note above), fork/star ratio 15.7% (median for new repos with ≥5000 stars is about 11%), suspect is empty. trending_days_any = 0. This is because GitHub trending only includes repos at least about 60 days old, which doesn't mean it wasn't popular.

- Trendshift (https://trendshift.io/repositories/319002): from 10-21 to 11-06, +4,878 stars, +880 forks, 28 merged PRs, and only 1 contributors.

- Strong engagement: 232 issues and 142 PRs for a desktop app that is 3 weeks old and maintained by one person. That's a real signal that people are installing it, using it, and requesting features (actual usage numbers unverified).

### 3. Where users came from (growth tactics)

| Channel | Evidence | Confirmed? |
|---|---|---|
| **The author's own YouTube channel (main driver)** | Channel @EeroAlvar with 30.7k subscribers and 18 videos. "Agent Harness for Learning Things" was published 10-23, with 467k views and 11k likes. The repo link is in the first line of the description. Stars began rising the day after release. | Yes |
| **Audience and trust built earlier** | "How I Use AI to Learn Things" (published around 2026-08-14) reached 1M views. Its companion repo `learn` ("My AI learning system from this video") went from 673 stars (08-29 on Wayback) to 2,970 (10-01), and is now around 4,046. "Bodybuilding for the Mind" got 444k, and "Pi Setup After 6 Months" got 216k. The author also has pi-config at 1,236 stars, among others, for about 11k stars in total. | Yes |
| Recognized by the Pi coding agent community | The author's early videos and repos centered on the pi coding agent (pi-subagents, pi-dictate, pi-observational-memory). | Yes (context) |
| "agent harness" buzz | YC ran a harness special in 2026 September (09-07). Greg Isenberg's "AGENT HARNESSES ARE THE NEW GPT WRAPPERS" (09-13.7M followers), and elvis/omarsar0 back-to-back "build your own harness" posts (x-radar corpus). | Yes (timing). How much it contributed to the star count is unverified. |
| Aggregator / SEO secondary spread | explainx.ai (10-25 "1.5k Stars in 9 Days"), reveneau (10-25), teckdeck (10-26), gipyeong-lee Korean blog (10-27), MDALatam Spanish LinkedIn (10-27), webconsulting.at, aifrontierpost (10-28), coderefer (10-29), chen.media (11-01 "Get a Free Claude Tutor"), arnavsheth.com (11-05). | Yes |
| Chinese X repost accounts | Search snippets show GitHubDaily on X ("Let AI explain a hard concept…") and Geek Lite on X. The x-radar local corpus didn't turn up learnharness, so the timing and engagement of the original posts are unverified. | Partial |
| Trendshift charts | 10-24 to 10-29 it reached the TypeScript top25 (peak #3) and stayed in the site-wide top25 for 3 days. | Yes |
| Hacker News | One post on 10-26 got 14 points and 0 comments. It didn't take off. | Yes (didn't work) |
| Ads / paid acquisition | None found. It's open source and free with no revenue, so there's no reason to spend on ads. | — |

**Bottom line** : the main driver was the author's own YouTube videos, plus an audience he had already proven with the "how to learn with AI" topic. Aggregator sites and Chinese repost accounts amplified it. HN and GitHub trending did not help.

### 4. Competitors

| Name | Where | Launched | Scale | How it differs from LearnHarness |
|---|---|---|---|---|
| HKUDS/DeepTutor | Open source | 2025-12-28 | 55.4k stars, +5,749 in the last 7 days (our repo) | HKU lab project: RAG + multi-agent + document-QA tutoring, Python web / CLI. Focus is on generating questions from your material and answering doubts, not on first probing the edges of your understanding and then reasoning step by step through a dependency graph. |
| Mr. Ranedeer AI Tutor | Open source (prompt) | 2023-03 | 29.5k stars, stalled | A single GPT-4 prompt, no state, no knowledge graph, no practice scheduling |
| amosblomqvist/learn (the author's own earlier version) | Open source | 2026-08 | About 4,046 stars | The author's personal config / skills shared as-is, which need an agent environment like pi to run; LearnHarness is a desktop app for ordinary people |
| hhyqhh/inno-agent | Open source | About 2026-05 | 1,381 stars | A personal learning agent also built on the Pi SDK (three-layer memory + IM channels), positioned as a 'learning partner', not step-by-step reasoning-style teaching |
| Miaotofu01/Study-Mate | Open source | About 2026-09 | 1,358 stars | Chinese 'AI learning partner' that plans routes and builds projects |
| ChatGPT Study Mode / Claude Learning mode / Gemini Guided Learning | Big-company features | Since 2025 | Massive user base | Built into a chat product, Socratic Q&A, but no explicit cross-session knowledge graph and no separate practice sessions; the prompt can't be modified and data lives in the cloud |
| Google Learn Your Way, NotebookLM | Big-company features | 2025 | Big | Rewrites material into personalized textbooks / podcasts; leans toward 'reading', with no diagnosis or reasoning steps |
| Akas333/TYSA(Teach Yourself Anything) | Open source | About 2026-10-26 | 21 stars | Copycat: the intro almost copies 'learning harness that finds where your understanding…', and it's also Electron + knowledge graph |

### 5. Copycats

- **Akas333/TYSA** : repo created about 11 days ago, 21 stars, intro and positioning directly mirror it ("a learning harness that finds where your understanding…", topics include harness and knowledge-graph).

- A few other small repos from the same period: BaronCyrus/dsh-aha-tutor (a private tutor for DeepSeek Harness, 27 stars) and smashburger-dev/socratic-workshop (a Socratic learning skill for Claude Code / Codex, 23 stars). None took off.

- The category term 'learning harness' is already being copied, but nobody has copied it with their own distribution channel.

### 6. Why it took off

1. **The distribution is the author's own** : an education / AI creator with 30k subscribers, whose single videos can pull hundreds of thousands to over a million views. The video went out 10-23, stars started climbing 10-24, and the timing lines up exactly; the first line of the video description is the repo link.

2. **He validated the demand once already** : "How I Use AI to Learn Things" got 1M views + the `learn` repo got 4,000 stars, showing that an audience for 'how to actually learn hard things with AI' already exists and had already converted into stars. This is the 'official version' for the same audience.

3. **He caught a timing buzzword** : since 2026 September, 'agent harness' has become a hot term on X and in YC circles. Calling a tutoring app 'Agent Harness for Learning' is instantly legible to developers, who are willing to click.

4. **Near-zero barrier for developers** : free, AGPL, local storage, bring-your-own key, and it **can directly use the Claude Code CLI / existing subscriptions** (chen.media's headline is literally "Get a Free Claude Tutor"). People who already pay for an LLM subscription don't need to pay again.

5. **It's real, so forks and issues are high** : installers for all three platforms (macOS signed and notarized), in-app updates, knowledge graph, thinking panel, dictation, background research; shipped to v0.4.1 with 28 merged PRs. Fork-to-star ratio is 15.7%, higher than comparable new repos.

### 7. Differentiation

- **Methodology** : it doesn't 'explain it to you'. It first measures where your understanding breaks, then reasons through a dependency graph step by step, fixes misconceptions before building on top, and then runs separate sessions of unnamed retrieval practice. This is diagnostic teaching + retrieval practice from cognitive science, a philosophy the author has repeated across several videos. Big-company study modes mostly stop at Socratic questioning.

- **The knowledge graph stores 'propositions + reasoning chains', not skill scores** , and can be reused across sessions.

- **Local-first, editable prompts, user-chosen models (including subscriptions)** , which suits developers and self-learners.

- **Are these differentiators why it took off?** Only half. The starting point and pace of the stars were almost entirely driven by the YouTube video. The differentiators give the video something to talk about ("AI explaining it once just makes you feel like you understand" is a pain point that resonates), and make viewers willing to install and open issues. The copycat with the same idea (TYSA) had no distribution and stalled at 21 stars.

### 8. Exposure assessment

**Circulating in small circles** : the YouTube video has 467k views, and about a dozen AI aggregator / SEO sites and Chinese, Korean and Spanish-language communities passed it around, but there's no mainstream media coverage; x-radar's local corpus has no big accounts discussing it; HN didn't pick it up. No revenue involved (free and open source).

### 9. What others can learn

- **The 'video + open-source companion tool' playbook** : first validate demand with a methodology video (1M views), then ship a companion personal-use repo (4K stars), then the formal product + a second video (5.5K stars and still climbing). Each step builds the audience for the next.

- **Applying a hot term to a non-developer scenario** (harness → learning) makes it legible to both sides.

- **Letting users plug in existing subscriptions (Claude Code CLI, etc.)** means 'free tutor' for developers, which is very easy to share.

- The README's first screen lays out the four steps Measure / Plan / Teach / Practice clearly, with one screenshot and one knowledge-graph diagram, so it's understood at a glance.

### 10. Still unclear

- Actual download / active user numbers: release downloads not checked (the GitHub API was rate-limited at the time).

- I haven't verified the post times or engagement for the Chinese-language X posts from GitHubDaily and Geek Lite, so I can't say how much they contributed to star counts on 10-27/28.

- The +1,125 in our repo log 10-28 and the +535 on Trendshift don't match; I treated this as a snapshot timing difference and didn't verify hour by hour.

- Whether the author will charge later (the README says that at most they'd charge for hosted keys / hosted services).

- No public research on learning outcomes (explainx also pointed this out).

### Sources

- GitHub README: [github.com/amosblomqvist/learnharness](https://github.com/amosblomqvist/learnharness) ; author profile [github.com/amosblomqvist?tab=repositories](https://github.com/amosblomqvist?tab=repositories)

- Our GitHub repo (agent-im server) repo_profile / history, 2026-11-06

- Trendshift:[trendshift.io/repositories/319002](https://trendshift.io/repositories/319002)

- YouTube: [www.youtube.com/watch?v=nfTDljFG2g0](https://www.youtube.com/watch?v=nfTDljFG2g0) (467,566 views, 2026-10-23); channel [www.youtube.com/@EeroAlvar/videos](https://www.youtube.com/@EeroAlvar/videos)

- "How I Use AI to Learn Things": [www.youtube.com/watch?v=kzcI5F4tGiU](https://www.youtube.com/watch?v=kzcI5F4tGiU) ; podwise 2026-08-14

- explainx:[explainx.ai/blog/learnharness-open-source-agent-harness-for-…](https://explainx.ai/blog/learnharness-open-source-agent-harness-for-learning-knowledge-graph-2026)

- reveneau:[reveneau.com/ainews/learnharness-agent-desktop-app-knowledge…](https://reveneau.com/ainews/learnharness-agent-desktop-app-knowledge-graph-1115-stars)

- chen.media:[chen.media/guides/get-a-free-claude-tutor-that-finds-exactly…](https://chen.media/guides/get-a-free-claude-tutor-that-finds-exactly-where-you-got-stuck)

- MDALatam LinkedIn (Spanish): [es.linkedin.com/posts/mdalatam_quieres-aprender-un-tema-dif%…](https://es.linkedin.com/posts/mdalatam_quieres-aprender-un-tema-dif%C3%ADcil-con-un-activity-7520898913696944129-G7hT)

- Our HN repo: story 50243281 (2026-10-26,14 points)

- x-radar corpus: posts about agent harness from YC 09-07, gregisenberg 09-13, omarsar0 09-11/12/14

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.