The free AI notetaker is genuinely free. Here is why I self-host anyway.

This is the write-up. It starts with the parts that hurt.

Concession one: Fathom's free tier beats mine on cost

Fathom's free plan is unlimited recordings, unlimited storage, and unlimited transcription in 38 languages, forever, for zero dollars, with zero setup (fathom.ai/pricing and Fathom's help center, both fetched 2026-07-24). The fine print is milder than you would expect: advanced AI summaries are capped at your first 5 calls each month, and features like AI action items and follow-up emails are Premium.

Backchannel is $0 in license terms. It is not $0 in practice. It costs a Docker host, optionally a GPU, and either LLM API spend on your own key or the accuracy hit of running local ONNX models. Anyone who tells you a self-hosted meeting assistant is "cheaper than Fathom" for a single user is lying to you. The cost argument only starts working at team scale -- Fathom Business is $25/user/month billed annually, so 20 seats is $6,000 a year -- or when your constraint is where the data lives rather than what the software costs.

Concession two: "no bot joins your call" stopped being special

For years the self-hosted pitch leaned on the visible recorder bot in the participant list. That era is ending. Fathom shipped bot-free capture in April 2026 (Mac only, Windows "coming soon"). Avoma's pricing page lists "Bot-less Native Recording" alongside bot capture. Granola and Otter already had non-bot desktop capture. The bot is still the loudest complaint in public reviews of Fireflies, tl;dv, and Read AI, but "no bot" is now a feature the incumbents advertise too. If an open-source project leads with it in 2026, it is following, not differentiating.

Concession three: real-time in-call AI stopped being novel -- in one week

Within a single week in July 2026, real-time meeting assistance went from rare to shipping at two of the biggest vendors. Otter announced Live Assist on 2026-07-21, positioned as a real-time AI coaching agent that joins any call -- available on Otter Enterprise only. Zoom took Sales Assist to general availability on 2026-07-22 inside Zoom Revenue Accelerator, at $66/user/month (Essentials) to $99.99/user/month (Premium), billed annually -- the only vendor in my entire research set that publishes a real price for real-time assistance. Avoma sells a Real-time Answer Assistant inside a $29-35/user/month add-on, which lands at roughly $53-74 per user per month all-in.

So I cannot claim novelty, and I will not. What I can describe is the mechanism, because it is different in a way that matters.

What the incumbents' live AI actually does

Zoom documents its system best, so I will use Zoom as the example. Per Zoom's own support documentation, an administrator pre-configures talking points for each topic category -- discovery, objection handling, pricing, competitor talking points -- attaches knowledge collections as reference, and rule-based live signals and timers drive when a card is displayed. It is a well-built retrieval system over content that was written before the call started.

Backchannel's objection handler does something else. Every 10 seconds it sends the trailing 90 seconds of the actual transcript to a low-latency model, with the meeting context, the speaker roster, your standing directives, and a do-not-repeat list. It generates two things: a response_now -- one or two conversational sentences you could say in the next ten seconds -- and a bigger_picture naming the underlying concern (cost justification, risk aversion, change fatigue, a missing stakeholder) and the strategic angle for later in the call. Nothing was authored in advance. It reasons over what was just said.

The honest counterpoint, and I want to state it rather than have a commenter state it for me: a curated battlecard is deterministic and enablement-approved. A generated response is neither. If your org needs every in-call prompt to have passed legal review, the prepared-content model is the correct one for you.

What the fine print says about "free" and "unlimited"

The category's pricing pages and help centers frequently describe two different products. A few examples, all from vendor-primary sources fetched 2026-07-24:

• tl;dv's free plan advertises unlimited recordings. The help center (updated 2025-11-24) adds: your first 10 meetings get automatic AI notes, after which only the first 10 minutes of each meeting does; you get 10 AI prompts and 10 AI reports -- and "these limits are for the lifetime of the account, they do not renew monthly." Free recordings are stored for 3 months, then gone.
• Read AI's free tier is 5 meeting transcripts per month with a 1-hour cap, and audio playback of your own meetings is an Enterprise-tier feature. Read AI also has a documented governance problem: Chapman University's IS&T published a notice on 2025-08-13 stating that "the use of Read AI is prohibited due to security, privacy, and institutional data risks," and UW-Madison's IT knowledge base documents its calendar auto-join behavior. Read AI disputes the framing; the institutional notices are citable either way.
• Otter's free tier is 300 minutes per month with a 30-minute cap per conversation and 3 file imports for the lifetime of the account.
• Fathom's free tier really is the good one -- unlimited recording, storage, and transcription. Per Fathom's privacy documentation, Fathom may use de-identified data derived from meeting content to train its in-house AI models, with an opt-out in account settings. That is the trade. The free tool is free partly because the meetings feed the model, and the control is opt-out, not opt-in.

None of this is scandalous. It is just the shape of the business model, and you should know it before you route every conversation you have through it.

What actually survives as a difference

After conceding cost, bot-free capture, and real-time novelty, here is what was left standing at the end of the research -- the rows where the answer for every mainstream tool was "no" and the answer for a self-hosted deployment was "yes":

• Where audio is processed. Fathom, Otter, Read AI, tl;dv, and Avoma all process audio in their cloud. Backchannel processes it on your server. Diarization -- who said what -- is always local (Silero VAD plus WeSpeaker ResNet152 speaker embeddings, both ONNX, both on your machine). Transcription is local ONNX Whisper/Parakeet if you choose, or your own API key if you choose.
• Whether a vendor trains on your conversations. With no vendor in the path, the question dissolves.
• Whether the live output is tactical content or a description. Live summaries and talk-time meters tell you what is happening. A generated next sentence is a different category of output, and in this tier nobody ships it at any price -- Otter and Avoma gate live assistance behind Enterprise or an add-on, and Fathom's and Read AI's live layer is summaries and behavioral metrics.
• Whether you can change the machine. Every Backchannel agent's model, prompt, and trigger interval is editable in the admin panel. You can also replay any recorded call through a different transcription model later and diff the results.
• Whether you can read the source. MIT license. For an IT security reviewer, source code is a different kind of answer than a questionnaire.

And the honest inverse list -- what the incumbents have that Backchannel does not: compliance attestations (SOC 2, ISO, HIPAA badges -- I have an architectural argument, not an audit), CRM sync, mobile apps, marketed language support, mature account and permission systems, and hundreds of thousands of companies of social proof. If you need those, you know who to buy from.

Who should use what

If you are one person who wants free meeting notes with zero setup, use Fathom. I mean that without irony; its free tier is the best deal in the category and my research did not find a catch big enough to change that sentence.

If your constraint is data location, if your review process prefers source code to vendor questionnaires, if you want live tactical help rather than a live summary, or if you are paying per-seat prices for a team -- that is the quadrant where self-hosting stops being a hobby and starts being the correct engineering decision.

The full sourced comparison tables, with access dates on every claim, are at backchannel.page/open-source-meeting-assistants/. The code is MIT and runs with docker compose up.