Content

Turn URLs, files, and text into media, knowledge, and documents.

One self-hosted engine. HomeTube client for YouTube — Content Studio for everything else — HomeTube browser extension — MCP server — CLI — Typed SDK — REST API underneath them all

Quick start · HomeTube · Coming from HomeTube? · All the clients · Connect an agent · Read the docs

Content is a self-hosted engine that turns supported sources into media, transcripts, summaries, translations, images, metadata, Markdown, PDF, and more. The backend analyzes each source, determines what it can actually produce, plans and runs the job, records its history, and delivers the results.

Everything else is a client of one public contract: HomeTube for the focused YouTube experience, Content Studio for general workflows, a Chromium extension for the tab you are already on, the content-mcp server for agents, a CLI for terminals and cron, a typed Python SDK for applications, the REST API for every other language, and Content Console to watch the engine work. Pick one or several — none of them is a layer the others have to pass through, and none of them holds business logic of its own.

HomeTube is the quickest way to see what that means: paste a YouTube URL, pick what you want out of it, and watch the files land in your library.

Quick start

Docker Compose is the only runtime prerequisite. Install from the published images — nothing to clone or build:

mkdir content && cd content
curl -fsSLO https://raw.githubusercontent.com/LatentNoise/content/main/deploy/docker-compose.yml
curl -fsSL -o .env https://raw.githubusercontent.com/LatentNoise/content/main/.env.example
docker compose up -d

The default .env starts the complete local stack:

Open Content Studio for the general workflow, or HomeTube if your source is YouTube. You can also skip both web apps entirely and drive the same engine from the MCP server, CLI, SDK, extension, or REST API.

Everything stays in the installation folder: ./data holds the database, jobs, and artifacts; finished files are delivered to ./playground/output. Set CONTENT_DELIVERY_DIR_HOST in .env to point at your own NAS or media library instead — its sub-folders then become the destination choices offered in the clients.

Update later with:

docker compose pull && docker compose up -d

git clone https://github.com/LatentNoise/content.git
cd content
cp .env.example .env
docker compose up -d --build

One source, many artifacts

The same engine serves media and document workflows:

                             ┌─ My Conference.mp4
                             ├─ My Conference - audio.opus
YouTube URL ──→ Content ─────┼─ My Conference - subtitles - en.srt
                             ├─ My Conference - subtitles - fr.srt
                             ├─ My Conference - transcript.txt
                             ├─ My Conference - summary.md
                             └─ My Conference - summary.pdf

Web page / text / .md file ──→ Content ──┬─ Article.txt
                                         ├─ Article - summary.md
                                         ├─ Article - translation.md
                                         └─ Article.pdf

Each branch starts with a declarative request. You describe the results; Content resolves what is possible, plans the work, runs the available tools, and records where every artifact came from. No yt-dlp flags, ffmpeg pipelines, transcription glue, or LLM orchestration in the client.

The clients

One engine, one contract, several independent front doors. Every client below speaks the same GenerationRequest, sees the same resolved capabilities, and produces the same artifacts — they differ in ergonomics, not in what they can ask for. HomeTube is entirely optional; so is every other row.

HomeTube — YouTube, focused

http://localhost:8501 · README

Paste a URL, choose the media and related artifacts, and follow the job into your library — the workflow demonstrated at the top of this page.

With HomeTube in Content, you can:

• download a video or playlist as video or audio, with quality, codec, container, audio-language, and subtitle choices;
• remove or mark sponsored segments with SponsorBlock, cut clips, use server-side cookie credentials, and embed subtitles, chapters, thumbnails, and metadata;
• give every artifact a readable name and deliver it into a filesystem library watched by Plex, Jellyfin, or Emby;
• ask the same source for a transcript, summary, thumbnail, or metadata — not just the downloaded media — when the required runners are available.

HomeTube is deliberately focused: it does not accept file, upload, or inline-text sources, and it does not expose the broader document workflow or PDF output. For those, use Content Studio, the API, CLI, or SDK. HomeTube is a Content client, not a layer every Content user must pass through.

It has no settings of its own. It reads the engine's configuration, so what changes HomeTube lives in the .env beside your docker-compose.yml: the delivery library whose sub-folders become the destination choices (CONTENT_DELIVERY_DIR_HOST), the languages pre-selected for audio and subtitles (CONTENT_LANGUAGE_PRIMARY, CONTENT_LANGUAGES_SECONDARIES, CONTENT_VO_FIRST, CONTENT_LANGUAGE_PRIMARY_INCLUDED_IN_SUBTITLES), and the server-side cookie file that unlocks age-restricted or members-only videos (CONTENT_CREDENTIALS — HomeTube only ever shows its id; the file never leaves the server). Nothing is ever pre-selected that the source does not offer, and no default is final.

HomeTube's README documents each variable and works the language rules through a concrete example — a Japanese talk, a French speaker, and exactly what ends up pre-checked. The full variable table lists everything the engine accepts.

Coming from standalone HomeTube?

HomeTube — the standalone Streamlit app — keeps running and stays maintained. Nothing breaks, and there is no deadline. Content is where the work continues, so this is a move you make when you are ready, not one you are forced into.

What comes across. The workflow you know: paste a URL, choose quality, codec, container, audio languages and subtitles; remove or mark sponsored segments with SponsorBlock; cut clips; unlock restricted videos with a server-side cookie file; embed subtitles, chapters, thumbnails and metadata; and deliver readable file names into a library watched by Plex, Jellyfin, or Emby. Playlists come across too — each entry is analyzed and planned on its own, named per item, and delivered as its own artifact.

What is new. The same source can also produce a transcript, a summary, a translation, generated thumbnails, metadata, Markdown, or PDF. And the engine is no longer reachable only from a web form: Studio, the Chromium extension, an MCP agent, the CLI, the SDK, and plain HTTP all ask for the same things.

What has not made the trip yet. HomeTube's playlist synchronization — the plan/apply diff that keeps a local folder in step with a playlist over time (rename detection, archiving or deleting removed videos, relocation after a naming or location change) — does not exist here. Content downloads a playlist; it does not yet re-sync one. It is roadmapped as part of M2. If playlist sync is your main use, keep standalone HomeTube for that job.

One practical note. Both HomeTube apps default to port 8501. Change one of them in your .env if you want to run the two side by side.

Content Studio — the whole contract, in a form

http://localhost:8502 · README

The general-purpose web app: several sources in one request, every output type the engine can resolve, with their options, preferences, and constraints. The form is capability-driven — it renders what /capabilities answers for your source, so a capability the engine gains appears without a UI release, and what cannot be produced is shown with the server's reason instead of silently disappearing.

Studio takes a file two ways, and the distinction matters when the engine runs on another machine:

• From this device — the bytes are sent to the engine and become an upload source (ADR 0020). Several files at once become several sources. Streamlit buffers the upload in the UI container, so the ceiling here is deliberately lower than the API's (200 MB by default, and the picker states the current value rather than letting you discover it by failing).
• On the server — a path the engine can already read, under a configured allowed input root. With this Compose setup that root is ./playground/input.

For genuinely large files, skip the browser: the SDK streams from disk with client.upload_file(path).

Chromium extension — the tab you are already on

README · latest release

Send the page you are watching to your engine without opening a UI. Manifest V3, Chromium only (Chrome, Brave, Edge, Vivaldi, Opera, Arc…), no build step — what the browser loads is exactly the files in the zip:

1. download content-browser-extension-chromium-v<version>.zip from the latest release and unzip it (Chromium loads a folder, not a zip);
2. open chrome://extensions, turn on Developer mode, Load unpacked, select the folder;
3. open a video and click the extension.

It normalizes the tab's URL (youtu.be, Shorts, and embeds become the canonical watch URL; a list= on a watch URL means that video, not the playlist), asks the engine what the source can produce and offers only that, prefills the name with the naming engine's own proposal, offers the library's existing folders plus new folder…, then follows the job and shows where each file landed. Every network call happens in the service worker — the engine sends no CORS headers by default, and that is what lets the extension work against a stock instance with nothing to configure (ADR 0016).

The popup's footer always names the engine it is talking to, so "where is this sending my video?" is answered on screen. Its README also carries an honest path-by-path verification table — what has actually been driven in a browser, and what has not.

MCP server — for agents and IDEs

README · content-mcp on PyPI

Content gives an MCP-compatible agent a controlled way to turn URLs into real artifacts, not just talk about them. The official content-mcp server exposes intention-level tools — not one per endpoint — to:

• analyze a URL and discover what it can produce;
• request one or several outputs and choose a library destination;
• monitor the job, cancel it, and report actionable failures;
• find the artifacts and their delivered paths;
• read small text artifacts inline while keeping large media out of the model context.

The agent never needs shell access, yt-dlp syntax, or backend internals.

Nothing to install. uvx fetches the server on first use and caches it, so the client owns the whole lifecycle (updating stays explicit — uvx --refresh, or pin content-mcp@x.y.z):

# Claude Code
claude mcp add content \
  --env CONTENT_API_URL=http://localhost:8010 \
  -- uvx content-mcp

For Claude Desktop, Cursor, and other MCP clients using the standard JSON shape:

{
  "mcpServers": {
    "content": {
      "command": "uvx",
      "args": ["content-mcp"],
      "env": { "CONTENT_API_URL": "http://localhost:8010" }
    }
  }
}

Prefer a pinned executable on your PATH — for an offline machine, or to control when the version changes? Install it and name it directly ("command": "content-mcp", or -- content-mcp for Claude Code):

uv tool install content-mcp      # or: pipx install content-mcp

Then ask naturally:

Analyze this lecture, save its audio into Talks, create a transcript and a structured summary if the available runners allow it, and tell me where every artifact landed.

The normal tool journey is get_config → analyze_source → generate → get_job → get_artifact. Downloads to the agent's own machine are bounded by a single variable (CONTENT_MCP_DOWNLOAD_DIR, ~/Downloads/Content by default): a destination pointing outside it is refused, not clamped — widening that is the operator's decision, not something a prompt can talk the server into.

What it can and cannot do, briefly — the MCP guide has the full three tables, and each row there was driven over stdio against a running engine rather than inferred:

CLI — terminals, scripts, and cron

README · content-cli on PyPI

uv tool install content-cli
export CONTENT_API_URL=http://nas.local:8010

content analyze "https://www.youtube.com/watch?v=…"
content video "https://…" --height 1080 --subs en,fr --watch
content audio "https://…" --format opus --playlist --watch
content submit request.json --watch          # a raw GenerationRequest, or -
content jobs ; content job <id> ; content artifacts <id>
content download <artifact_id> -o out.mkv

--watch follows the job's event stream rather than re-asking for its status, and the exit code carries the outcome so a script can chain on it (ADR 0021): 0 succeeded, 2 partially succeeded, 1 failed or cancelled. 2 is deliberately distinct — a playlist that yielded five videos of six is not a failure, but a script that treats it as a success will quietly move on with missing files.

Python SDK — typed, sync and async

README · content-sdk on PyPI

The one official API client: the CLI, the MCP server, and the web apps all speak through it, so the engine's rules are never duplicated. Python 3.11+, httpx and pydantic only.

from content_sdk import ContentClient, outputs

with ContentClient("http://localhost:8010") as client:
    analysis = client.analyze(outputs.url_source("https://www.youtube.com/watch?v=…"))
    job = client.generate(analysis.id, [outputs.audio_output()])
    job.wait()
    for artifact in job.artifacts:
        print(artifact.display_filename, artifact.delivered_path)

AsyncContentClient mirrors the whole surface. Analyses are addressable, so every call accepts either an analysis_id or inline sources (ADR 0014); client.upload_file(path) streams a local file to the engine and hands back a ready-to-use source; and every non-2xx becomes a typed exception carrying the stable error codes (NotFound, Gone, ValidationError).

REST API — every other language

Public contract · Swagger at http://localhost:8010/docs

/api/v1 is the contract every client above is built on, and it is stable enough to build a ninth client against: requests describe desired outputs rather than technical operations, error codes are machine identifiers that do not change (only their human messages do), and "valid but not implemented" is a different, explicit answer from "invalid". Reserved fields are refused rather than silently ignored. The V1 API has no built-in authentication — see deployment and security.

Content Console — observe and pilot the engine

http://localhost:8503 · README

The operations console: strictly observability and control, and deliberately not a way to create downloads (that is Studio and HomeTube).

• Overview — version, cache, concurrency, analysis TTL, a live jobs pulse, the installed runners with their availability, storage paths.
• Environment — every CONTENT_* variable with its effective value, whether it was set or fell back to a default, and a one-line description. Secrets are never shown — only their presence and length.
• Jobs — a filterable list plus full detail: steps, the submitted GenerationRequest, artifacts with their provenance, ordered events, per-step logs, cancel and retry, with an opt-in 5-second auto-refresh that polls only while jobs are in flight.
• Storage & cache, and a contract & API tab with the schemas, Swagger, ReDoc, and a raw API tester.

What Content can produce

Content resolves availability for each analyzed resource and installed runner. Clients show what can actually be produced instead of presenting a static list and failing halfway through a job.

Media acquisition works with YouTube and the sites supported by yt-dlp. Playlists can produce artifacts for each member in a single traceable job. Finished files can be delivered directly into any mounted filesystem library, including folders watched by Plex, Jellyfin, or Emby.

Local and optional AI runners

Content has no mandatory cloud dependency. AI-backed outputs activate when a compatible runner is available:

• Summaries, translations, and derived chapters: connect a local Ollama instance, or explicitly configure an Anthropic or OpenAI API key. Cloud runners can be excluded per request.
• Transcription from audio: run a speech service next to the engine — docker compose --profile speech up -d and CONTENT_SPEECH_URL=http://speech:8000, or speech.enabled: true in the Helm chart. It is speaches, behind the OpenAI audio API, and the same container also does text-to-speech. Audio is only ever sent to a service on a private network. Transcripts from existing subtitles do not need it.

Unavailable optional runners do not make Content unhealthy. The affected capability is reported as unavailable, and an impossible request is refused before the job starts. See deployment and configuration for the full inventory.

Why build on Content?

• Declare intent, not tooling. Requests describe outputs; yt-dlp, ffmpeg, Whisper, LLMs, and PDF renderers remain replaceable implementation details.
• Self-hosted and local-first. Run it on a workstation, NAS, or homelab. There is no account, telemetry, or mandatory cloud service.
• One public contract. Studio, HomeTube, MCP, the CLI, SDK, extension, and REST API converge on the same domain instead of drifting into parallel feature sets.
• Human files, not pipeline debris. The engine names every artifact and can place it directly in the library you already use.
• Observable work. Jobs expose states, ordered events, progress, logs, artifacts, provenance, cancellation, and retry.
• Honest capabilities. “Valid but unsupported,” “unavailable on this source,” and “broken” are different answers, reported before or during the correct stage.

How it works

The public request describes desired outputs. It does not prescribe technical operations:

Sources → Analysis → Capabilities → Execution plan → Job → Artifacts

Content Backend (REST /api/v1)   ← domain, planning, execution, persistence
        │
   Content Python SDK            ← official Python API client
    │        │        │
   CLI      MCP     Web apps     ← thin clients of the public contract

   Browser extension             ← JavaScript client of /api/v1

The engine analyzes source facts, resolves feasible capabilities against the installed implementations, builds a deterministic execution plan, and runs it as an observable job. The database is the source of truth; yt-dlp, ffmpeg, transcription, LLM, and PDF implementations stay behind dedicated boundaries.

The domain keeps user intent (GenerationRequest), resolved strategy (ExecutionPlan), execution (Job), and concrete results (Artifact) separate. For the full design, see the architecture overview, public contract, domain model, and architecture decisions.

Deployment and security

Content targets single-host Linux/macOS deployments on amd64 or arm64. Docker Compose runs the API and embedded worker together, with the web apps as separate API clients. SQLite and artifact storage persist locally; no Redis, Celery, Kubernetes, or cloud infrastructure is required.

The commented .env.example covers UI selection, image versions, ports, storage and delivery, language preferences, server-side cookie credentials, optional LLM/STT runners, CORS, and notifications. No release check makes an outbound request unless you configure one.

The V1 API has no built-in authentication: anyone who can reach it can submit jobs, upload files and read every artifact. Keep it on a trusted network, or put an authenticating reverse proxy in front of any externally reachable instance — and do not publish the port. What an attacker who reaches the API can and cannot do is written out in docs/operations/threat-model.md. See deployment for configuration, health checks, data layout, authenticated sources, and production guidance, and ADR 0024 for why that is the V1 answer.

Development

The repository is a monorepo containing the backend, web applications, CLI, MCP server, Chromium extension, and Python SDK. The root Makefile is the main entry point:

make install       # create the development environment and install packages
make validate      # formatting, linting, and hermetic tests: the official gate
make test-ui       # hermetic Streamlit application tests
make validate-all  # official gate plus UI and opt-in external-tool tests

The main test suite does not require the Internet. Read the repository architecture and validation guide before changing the code. Releases follow the release runbook.

Documentation

docs/README.md is the complete documentation map. Useful starting points:

• Product vision
• Public contract and domain model
• Architecture and invariants
• Storage and artifact delivery
• Deployment and operations

Licence

Content is source-available under the Functional Source License, Version 1.1, ALv2 Future License (FSL-1.1-ALv2).

In plain words: you may read, audit, modify and run Content — at home, for yourself, or inside your organisation. The one thing you may not do is offer it, or something substantially similar built from it, as a competing commercial product or service.

Every version becomes Apache 2.0 two years after it is published. That promise is in the licence itself and cannot be taken back.

Copyright 2026 Yann Orieult. "Content" and "Latent" are names of the project and of its author's work; a fork must not use them for its own product.

Versions up to and including 0.8.4 were released under AGPL-3.0-or-later and remain available under it, permanently. See COMMERCIAL.md for the full summary; no separate commercial offering is currently available.

Governance

Content is developed and maintained by Yann Orieult under a single-maintainer model. Bug reports, ideas, and design feedback are welcome through issues. Code contributions are not accepted—unsolicited pull requests are closed without review; forking is the intended path for independent changes. See CONTRIBUTING.md and GOVERNANCE.md.

Report security issues privately as described in SECURITY.md.

© 2026 Yann Orieult.