Adding an agent to a team can quietly add a second stack of subscriptions: meeting notes, task tracking, memory, model access, voice, web data, and chat. In this episode of The Next New Thing, Andrew Warner tours seven open-source projects that cover those jobs. His companion collection puts the tools and trade-offs in one place. The useful question is not whether all seven are free to download; it is which specific workflow each one can replace without moving cost or risk somewhere less visible.
Start with the job, not the logo. FluidVoice handles local dictation and, in its new FluidMeet feature, meeting transcripts. Paperclip coordinates agent work. Basic Memory keeps knowledge in Markdown. FreeLLMAPI routes across free model tiers. VoiceStudio makes voice media, Scrapling gathers web data, and Buzz provides agent-first team chat. None is automatically a feature-for-feature replacement for a hosted suite. Hardware, model rights, provider limits, setup, and maintenance still count.
Watch Andrew's Seven-Tool Tour
Credit: The demonstrations and companion deck are by Andrew Warner / The Next New Thing. The Zapier segment at 09:27 is a sponsored aside, not one of the seven repositories. Repository links below go to the projects themselves; feature status can change after publication.
Choose by the Job You Need Done
The episode groups several different kinds of software under “agent tools.” This quick map separates the work before the deeper notes.
| Job | Candidate | First check |
|---|---|---|
| Private speech and meetings | FluidVoice / FluidMeet | Mac support, recording consent, transcript quality |
| Assign and govern agent work | Paperclip | Agent costs, approvals, local operation |
| Keep reusable agent knowledge | Basic Memory | Markdown ownership and sync boundary |
| Experiment with model routing | FreeLLMAPI | Rate limits and non-production use |
| Create voice and dubbed media | VoiceStudio | Hardware, model licenses, speaker consent |
| Extract authorized web data | Scrapling | Site terms, robots.txt, data quality |
| Chat with humans and agents | Buzz | Self-hosting and unfinished features |
Speech, Meetings, and Voice Production
FluidVoice + FluidMeet
Where it helps
FluidVoice began as local dictation. Its v1.6.11 release introduces FluidMeet: local meeting recording, live captions, speaker labels, playback, transcripts, and a local API that can retrieve meeting records. This is a practical way to give an authorized local agent a transcript without automatically sending every call to a cloud note service.
What to verify
Andrew demonstrates the Mac workflow and points to Windows early access; do not assume cross-platform parity yet. The release does not make every feature of Granola available: check summary quality and sharing separately. Core software is open source, while Fluid Intelligence enhancement uses a separate private model. Obtain meeting consent and test transcription on your real microphone and accents.
VoiceStudio
Where it helps
VoiceStudio puts text-to-speech, voice design, voice cloning, audiobooks, and dubbing into a desktop workflow that can use local models. Andrew's example is producing an agent's narration or translating existing training material without paying per hosted generation.
What to verify
This is a production tool, not a blanket “free ElevenLabs.” Local inference needs suitable hardware and time. The application license, each model's license, and the rights to an input voice are separate questions; the repo explicitly warns that a software license does not grant rights to every model. Clone only with the speaker's permission and have a fluent reviewer check dubbed output.
Agent Teams and Memory
Paperclip
Where it helps
Paperclip treats agents and humans as workers with roles, task ownership, budgets, and oversight. Andrew shows the org chart and task handoffs rather than just another kanban board. It is worth testing when several coding or research agents need shared priorities and a person must review the work.
What to verify
It does not erase Claude, Codex, or other model costs. The open-source app can run locally, while hosted Paperclip Cloud is still presented as a waitlist in the project documentation. Compare your needed Jira features and audit rules; do not migrate the team's only task history on the strength of an agent demo.
Basic Memory
Where it helps
Basic Memory stores notes as ordinary Markdown files and exposes them to compatible assistants through MCP. Linked notes and search give an agent durable context across sessions while leaving the underlying files inspectable and portable. The local setup is the relevant alternative when your team mainly pays for a note tool to feed agents.
What to verify
A Markdown memory layer is not all of Notion: databases, permissions, publishing, and rich collaboration are separate needs. Decide which notes an agent may read or edit, how conflicts and backups work, and whether you want the optional hosted offering rather than only local files.
More on Paperclip: an Agent-Team Walkthrough
For a longer look at Paperclip beyond Andrew's roundup, NetworkChuck demonstrates his own AI IT department. It is a separate creator's setup, not evidence that every Paperclip deployment behaves the same way.
Model Access and Web Data
FreeLLMAPI
Where it helps
FreeLLMAPI routes requests across model providers' free tiers and presents a compatible endpoint for experiments. It can help a developer try a small classifier, compare outputs, or prototype a low-volume agent without committing every test to a paid API.
What to verify
“Free” means provider allowances with changing quotas and eligibility, not unlimited model access. The project's README explicitly positions the router for personal experimentation, not production use. It does not make paid Claude models free. Track rate limits, latency, quality, privacy terms, and fallback behavior before letting a business workflow depend on it.
Scrapling
Where it helps
Scrapling gives Python developers fetching, extraction, crawling, and agent-friendly output for web research they are allowed to perform. Its adaptive selectors can make a documented extraction workflow less fragile when a page layout changes.
What to verify
This is not permission to scrape any site or evade its controls. The project tells users to respect site terms and robots.txt and comply with privacy law. Check source rights, request volume, provenance, and data freshness. For a business service, prefer authorized APIs or exports whenever they meet the need.
Agent-First Team Chat
Where it helps
Buzz gives agents their own identities in team conversations. Its current repo lists channels, threads, DMs, search, media, audit logs, a desktop client, and a CLI/agent connection path. Andrew's strongest point is the shared conversation: people can see and redirect an agent's work in context.
What to verify
Do not call it a finished Slack replacement. The repo marks mobile clients and workflow approvals as still being wired up, and self-hosting brings server, database, storage, backup, and access-control work. Pilot one internal channel before moving operational communications.
“Free on GitHub” Is a Starting Price, Not a Total Cost
Andrew's companion deck gives subscription examples to make the opportunity concrete. Treat them as illustrations, not guaranteed savings: plans and usage tiers vary, and a self-hosted project is not always equivalent to the product beside it. The correct comparison includes the subscription you can actually cancel and local hardware, hosting, provider usage, setup time, updates, backups, security, and human review.
There is also a difference between open-source application code and all of the components it uses. FluidVoice's optional enhancement model is separately distributed; VoiceStudio's app and voice models can have different terms; FreeLLMAPI relies on outside providers. Paperclip and Buzz may be free to download while connected agent inference and operations are not. Check the exact license and deployment configuration for each intended client use.
A Sensible First Pilot
- Pick one paid job. For example, move a small set of consented meeting transcripts into FluidMeet, or let Paperclip coordinate a single internal agent task. Keep your current service running during the test.
- Write a baseline. Record current spend, time, error rate, privacy constraints, and the feature that cannot break.
- Test on ordinary and difficult cases. Use real accents for transcripts, a failed handoff for task tracking, or changing page layouts for permitted scraping.
- Measure total cost and reversibility. Include installation, maintenance, model calls, hardware, backups, and staff time. Export the data once before deciding to switch.
If the replacement is only cheaper when nobody counts maintenance, it is not cheaper. If it gives your team control that the paid tool cannot, that may still justify the work, but name the benefit accurately.
Jump to the Episode
| Time | Segment | Time | Segment |
|---|---|---|---|
| 00:00 | Why agent subscriptions add up | 00:27 | FluidVoice and FluidMeet |
| 03:09 | Paperclip | 06:18 | Basic Memory |
| 09:27 | Zapier sponsor segment | 10:30 | FreeLLMAPI |
| 14:42 | VoiceStudio | 18:27 | Scrapling |
| 21:00 | Buzz | 24:27 | Choosing agent-first tools |
Business Idea: A Measured Agent-Stack Cost Audit
A narrow service opportunity is to help one small team test whether one recurring agent expense is worth replacing. Sell the measured decision and migration support, not a promise that open source makes every bill disappear. The prompt below turns the episode's seven tools into a client-specific pilot with a stop rule.
Test one agent-stack replacement
A seven-day pilot with quality, cost, rights, and rollback checks.
Sources and Useful Links
Creator: Andrew Warner's full episode and companion collection. Original projects: FluidVoice (FluidMeet release), Paperclip, Basic Memory, FreeLLMAPI, VoiceStudio, Scrapling, and Buzz. Project updates on X: FluidVoice and Scrapling. Check the repositories for current support and license details before adopting any tool.