Replacing ten subscriptions sounds satisfying. But Andrew Warner's tour for The Next New Thing is more useful as a map of ten jobs: speech, private chat, model routing, web research, image creation, music, short clips, app building, computer-using agents, and upscaling. One open repo may replace the part you actually use, even when it cannot replace the whole paid service.
Start with your most repeated paid task, then test one repo against the current product. Open-source software can reduce subscription costs and increase control, but it may add model or API charges, GPU requirements, maintenance, and review work. Local is a configuration, not a promise about every feature. Keep the paid tool if the open alternative fails the quality or total-cost test.
Watch Andrew Warner's Ten-Tool Tour
Credit: Andrew Warner and The Next New Thing for the video and supplied transcript. The video mentions Zapier as a presenter. Andrew's Kill Your AI Subscriptions collection is the visual companion to this article. Project features and caveats below are cross-checked against the linked repositories; neither the video nor the collection is a substitute for current install and license documentation.
Choose by the Job You Need Done
| Job | Tool and repo | Paid-tool comparison | Check before switching |
|---|---|---|---|
| Speech | VoiceStudio | ElevenLabs-style voice workflows | Voice consent, model rights, hardware |
| AI chat | Odysseus | ChatGPT-style interface | Docker, model choice, data storage |
| Model access | FreeLLMAPI | Paid coding-model usage | Provider quotas and usage terms |
| Research | Vane | Perplexity-style search | Search and model providers |
| Images | InvokeAI | Hosted image studios | GPU and individual model licenses |
| Music | ACE-Step 1.5 | Suno-style generation | Output review and music rights |
| Short clips | OpenShorts | Opus Clip-style repurposing | Docker, AI API, edit quality |
| App building | Dyad | Lovable-style builder | Core/Pro licenses and model costs |
| Agent tasks | OpenManus | Manus-style workflows | Tool permissions and model setup |
| Upscaling | Upscayl | Topaz-style photo enlargement | Supported GPU and artifact review |
These comparisons describe overlapping jobs, not feature parity. Use the linked project pages for the latest supported platforms and install steps.
1-4: Voice, Chat, Coding Models, and Research
1. VoiceStudio: Local Speech With a Consent Check
At 00:27, Andrew shows VoiceStudio for speech generation, voice cloning, and dubbing. A local workflow can avoid per-clip hosted billing. It also makes you responsible for the compute and the voices. The repository says its app license is AGPL-3.0, while downloadable models and their outputs can carry separate terms. Only clone a voice with permission; test your intended commercial use against the exact engine you choose.
2. Odysseus: A Chat Interface You Operate
At 02:42, Odysseus is the private-assistant candidate. Its Docker-based setup can connect to local models and support memory and research workflows. The key question is where each request goes: a local model changes the privacy equation, but any configured external model or search provider may still see data. Test one thread, inspect storage and permissions, and compare answer quality before moving a whole knowledge base.
3. FreeLLMAPI: Failover Across Free-Tier Providers
At 05:51, FreeLLMAPI brings multiple provider keys behind one endpoint and can fail over when one tier is exhausted. That is convenient for experimentation, not an unlimited free coding plan. The repository explicitly frames aggregate free-provider usage as personal experimentation and directs users to each provider's terms and quotas. Do not design a production customer service around free credits that can change or disappear.
4. Vane: Self-Hosted, Source-Linked Research
At 09:27, Vane covers a Perplexity-like loop: search the web, gather pages, answer, and cite sources. Self-hosting lets you control the application, but web searches still contact external services and a remote model will receive the material you send it. Check each citation against the page and keep the instance access-controlled, especially if you attach internal research.
5-7 and 10: Create, Clip, and Upscale Media
5. InvokeAI: An Image Workbench
At 11:51, Andrew opens InvokeAI. It provides a local web interface with a canvas, inpainting, and repeatable node workflows for supported image models. The software has an Apache-2.0 core with additional notices in the repo; the specific image weights you download may have different licenses. A capable GPU and storage can matter more than the absence of a monthly software bill.
6. ACE-Step 1.5: Generate Music Locally
At 13:57, ACE-Step 1.5 is the Suno-style experiment: describe a track, then generate and iterate with a model on your own machine. The project publishes its code and model information on GitHub and warns about fake sites claiming to be official. Follow the repo's install links, not an unrelated download page. Review generated music for quality, rights, and resemblance before publishing or selling it.
7. OpenShorts: Repurpose Long Video
At 19:30, OpenShorts offers moment selection, vertical framing, subtitles, and optional dubbing. Its self-hosted core is MIT-licensed, but the current setup documents Docker and an AI/API configuration for the intelligent features; optional services may charge separately. An editor should still reject bad crops, wrong captions, and clips that remove necessary context. Test on one video and count publishable clips, not generated clips.
10. Upscayl: Enlarge Images in Batches
At 29:06, Upscayl is a straightforward first local test: enlarge a small set of images and compare faces, fine text, and edges against your paid upscaler. The desktop app is AGPL-3.0 and needs supported graphics hardware for local processing; the separate cloud offering is optional. AI upscaling invents plausible detail, so do not use it as forensic recovery of information that was never in the source.
8-9: Build Apps and Run Agents
8. Dyad: Build an App From a Brief
At 23:51, Dyad turns a plain-English brief into editable application code on your computer. It can use a model you provide, so code generation may still incur API cost. Check the repo's mixed licensing: the core is Apache-2.0, while src/pro uses Fair Source License 1.1. Treat the generated app as a prototype until someone reviews authentication, data handling, dependencies, and deployment.
9. OpenManus: An Agent Framework, Not a Finished Colleague
At 26:15, OpenManus is the broadest experiment: combine a model with tools for research, coding, and file tasks. Its code is MIT-licensed, but it does not include free unlimited inference. Give it a sandbox, limited credentials, and a task with an inspectable result. Require approval for external messages, purchases, file deletion, or production changes.
What “Free” Does Not Cover
- Compute: a capable GPU, electricity, disk space, and time spent installing or updating models.
- External services: model APIs, search providers, transcription, dubbing, hosted deployments, and publishing integrations may still bill you.
- Rights: a permissive app license does not override a downloaded model's terms, voice consent, media copyright, or mixed-license components.
- Operations: backups, access control, patches, failed jobs, and human review become your responsibility when you self-host.
For a fair test, record total cost per accepted result and turnaround time, not just the subscription line item. Keep a hybrid stack when a local tool handles routine work but a paid product is more reliable for difficult cases.
Explore the Companion Collection
Kill Your AI Subscriptions organizes Andrew's ten picks as a navigable visual collection with feature notes and direct repository links. Use it as a browsing guide; verify current setup, pricing, and license terms at each repo. Its bonus section also points to Affaan Mustafa's ECC, a separate open-source workflow toolkit for planning, building, testing, and reaching users. We covered the ECC interview and its actual outcomes here.
Bonus video from the companion collection: Andrew Warner with ECC creator Affaan Mustafa. It is an extra workflow example, not an eleventh subscription replacement in the main episode. The collection did not provide a verified original X post for each repo, so the project repositories are the source links here.
Turn One Replacement Into a Business Test
There may be a service in helping a specific buyer migrate one expensive, repetitive workflow and proving the result. This copyable prompt uses the tools above to explore that idea without assuming every open repo is free to operate or suitable for client data.
Find a paid workflow worth replacing
Three buyer-specific ideas, then one seven-day pilot.
Video Chapters
00:00 overview · 00:27 VoiceStudio · 02:42 Odysseus · 05:51 FreeLLMAPI · 09:27 Vane · 11:51 InvokeAI · 13:57 ACE-Step · 19:30 OpenShorts · 23:51 Dyad · 26:15 OpenManus · 29:06 Upscayl.
Source Map
- Creator sources: Andrew Warner's video, its supplied transcript, and the visual companion collection.
- Voice and knowledge: VoiceStudio, Odysseus, FreeLLMAPI, Vane.
- Media: InvokeAI, ACE-Step 1.5, OpenShorts, Upscayl.
- Apps and agents: Dyad and OpenManus. The bonus ECC repo and interview are separate from the main ten.