SourceHarbor

See It Fast

If the README is the front door, this page is the shop window.

This page is for fast evaluation. It is not a hosted demo, cloud sandbox, or one-click trial.

The goal here is simple:

1. show what SourceHarbor looks like
2. show what comes out of it
3. let you decide whether it is worth a deeper evaluation

If you like what you see here, the next step is run it locally, not “open the live app.”

The 20-Second Mental Model

SourceHarbor is not just a summarizer.

It is a full intake-to-digest loop:

• sources come in from YouTube, Bilibili, and RSS
• a job-backed pipeline processes each item
• operators read the result in a digest flow
• agents reuse the same evidence through API and MCP

The source story is intentionally uneven on purpose:

• YouTube and Bilibili are the strongest supported intake templates today
• RSSHub and generic RSS are real substrate paths, but they remain more generalized than the strongest video-first flows

The Three Surfaces That Matter First

1. Command Center

This is the operator home base.

What you should picture:

• subscription count
• discovered videos
• queued and failed jobs
• one place to trigger intake and inspect recent activity

Why it matters:

• it turns the repo from “a bunch of scripts” into a usable operating surface

2. Timeline

This is the reading surface.

Representative current feed shape:

• title: AI Weekly
• source label: YouTube · Tech Channel
• category label: Tech
• body path: digest markdown plus artifact metadata

Why it matters:

• the output is meant to be read, not just stored

3. Job Trace

This is the evidence surface.

What you inspect here:

• job_id
• status and pipeline final status
• step summary
• retry count
• artifact references

Why it matters:

• when something fails, you can debug with receipts instead of guesswork

What The Result Looks Like

SourceHarbor’s digest artifact template already tells the story of the output shape:

# <title>

> Source: [Original video](<source_url>)
> Platform: <platform> | Video ID: <video_uid> | Generated at: <generated_at>

## One-Minute Summary
<tldr>

## What This Covers
<summary>

## Key Takeaways
<highlights>

That is the key idea:

• not just transcript text
• not just one summary blob
• a reusable artifact with traceable structure

The 60-Second Evaluation Path

If you want confidence without booting the full stack yet:

1. Read README.md for the public story.
2. Read proof.md for the evidence ladder.
3. Read starter-packs/README.md if you want the public CLI / SDK / Codex / Claude Code starter surface.
4. Read docs/compat/openclaw.md if you specifically care about the new first-cut OpenClaw starter pack and its still-honest boundary.
5. Read samples/README.md if you want the clearly labeled sample corpus path.
6. Read architecture.md if you want the system map.

If you want a real local run after that, go to start-here.md.

Why This Attracts Builders

If you are evaluating whether this repo is worth starring, forking, or maintaining, this is the shortest honest filter:

You care about…	SourceHarbor answer
Codex / Claude Code fit	already exposed through MCP + HTTP API, with real Search / Ask / Job Trace surfaces behind it
AI product truth instead of AI vibes	proof, runtime truth, and project status all explain what is shipped, what is gated, and what is still a bet
A repo that feels like a product, not a pile of scripts	command center, digest feed, job trace, watchlists, trends, bundles, and sample playground all exist as coherent front doors
A contribution surface that is understandable	builder docs, compare docs, see-it-fast, and public truth surfaces reduce the amount of archaeology required before contributing

The honest lure is not “AI magic.” It is that SourceHarbor already gives builders:

• a Codex / Claude Code-friendly MCP and HTTP API surface
• a repo-local CLI substrate through ./bin/sourceharbor help when they want one discoverable command surface
• a public packaged CLI and public TypeScript SDK when they want versionable install/use examples without copying internal code
• a proof-first story that names external gates instead of hiding them
• a compounder layer worth revisiting when you care about watchlists, trends, and evidence bundles

This site is open source. Improve this page.