Every subscription price in this shortlist is zero except one. SillyTavern, Ollama, LM Studio and Open WebUI each record $0 a month and a $0 annual equivalent; HammerAI records $9 and $9. Verified dates: 2026-09-14 for SillyTavern, 2026-09-29 for HammerAI and Open WebUI, 2026-09-30 for Ollama and LM Studio. A column where four of five are identical cannot rank them, so the ranking comes from what it does not hold — which layer of the stack each one is, and what you have to supply yourself.
The rules that follow. If you want a companion on your own hardware and will supply the model, take SillyTavern as the interface and put Ollama or LM Studio underneath it: the runtime answers which weights, the frontend answers whether it remembers. If you want the same with a browser interface and no terminal, Open WebUI lists memory, voice and group chat together. If you want a free cloud allowance and a phone app in one product, HammerAI is the only entry with both, and its $9 buys volume rather than access.
One correction to the premise. Local running is not the mainstream of this category: 11 of the 154 SFW listings, about one in fourteen, against 145 with a free tier. It draws outsized attention not through adoption but because it is the only axis that changes who holds the conversation.
Four zeros and one nine, nothing in between
The price column, monthly then annual equivalent: SillyTavern $0 and $0; Ollama $0 and $0; LM Studio $0 and $0; Open WebUI $0 and $0; HammerAI $9 and $9.
For the four zeros the note behind the number matters more than the number. SillyTavern is free and open source under AGPL-3.0 and, in its own words, you pay only for whichever model API you connect (2026-09-14). Ollama is free and open-source, MIT (2026-09-30). LM Studio is free for personal use, with team licensing for work use (2026-09-30). Open WebUI is free and open-source, self-hosted via Docker or pip, with an enterprise tier that consumer use does not need (2026-09-29).
HammerAI is the exception and the arithmetic is short. Nine dollars a month against a nine-dollar annual equivalent means twelve monthly payments come to $108, the annual price is $108, and committing for a year saves nothing. No annual trap here, because there is no annual discount. Its free tier is 600 cloud messages a day plus unlimited chat on locally run models, no card required; Pro tiers at $9, $18 and $35 buy larger cloud limits. The paid decision is about volume, not access — all 600 used daily for thirty days would be 18,000 messages.
The price column stops where the bill starts
A vendor can only charge for what it delivers. Four of these five deliver software and stop; the model arrives from somewhere else, and that somewhere is where the money goes. Two routes exist, and only two.
The first is an API key: keep the free interface, pay a provider per token — recorded nowhere in the price table, since none of these vendors sells tokens. The second is a machine: buy, or already own, hardware that runs open weights, and the recurring cost becomes electricity against capital already spent. Ollama and LM Studio exist to make that second route work — one command to run open models, or a GUI over local models, per their own notes.
So a $0 stack is not automatically cheaper than $9: a light month on a metered key can land under nine dollars, a heavy month on the same key can exceed it several times over. The directory records no token prices, so this comparison cannot be finished from outside — the missing line item is your own usage. Sorting by what you supply produces the whole decision; by price, nothing.
Privacy claims are conditional in the same way. SillyTavern's note reads fully self-hosted, connect a local model and nothing leaves your machine; LM Studio's reads that chats and models stay on your machine. Both are true, both are conditional: point either frontend at a cloud model and the interface stays local while the conversation travels. Self-hosting protects the shell, not the payload, unless the weights are local too.
Memory is a frontend feature, not a runtime one
Read the capability rows and the five split in two.
Ollama and LM Studio are runtimes. Both list custom characters and then nothing else: no memory, no voice, no images, no group chat. They are not failing at companionship; they were never in that business. Their job is serving a model over a local port.
SillyTavern, Open WebUI and HammerAI are frontends, and all three list memory, voice, images, custom characters and group chat. The differences among them are narrow: SillyTavern adds voice calls, which the other two do not list.
That split explains the commonest disappointment in local setups. Someone installs a runtime, opens a raw model, notices it forgets yesterday's conversation, and concludes that local companions have no memory. Memory was never in the runtime. It sits in the layer above, which also decides whether a persona persists, whether lorebooks load, and whether a group shares a scene. Across the directory, 110 of 243 listings record memory; among these five, exactly the three frontends do.
Voice follows the same logic. Twenty-five of the 243 listings record voice calls, and here only SillyTavern claims them. If spoken conversation rather than typed chat is the requirement, one of these five is the shortlist.
The licence line decides the office question
Three are open source and the licences are not interchangeable once money is involved. SillyTavern is AGPL-3.0, the copyleft option: the obligation appears when you modify it and serve it to others, as an internal deployment does. Ollama is MIT, the permissive option and the least to think about. Open WebUI is free and open-source for consumers, with an enterprise tier sold separately.
The other two handle work use in their price notes rather than a licence. LM Studio is free for personal use and names team licensing for work — the one entry whose pricing already draws the consumer line. HammerAI is described as open-source with desktop apps and sells Pro tiers by cloud volume rather than by seat.
The rule: for personal use all five are available and the licence question is academic. For a company deployment, two notes already point at a commercial path, and SillyTavern's AGPL is the one to read before you modify and host it for colleagues. Where a vendor publishes no term, this page leaves it blank.
Desktop is the entry fee, and one of five has a phone
Platforms as listed: SillyTavern desktop and web; Ollama desktop; LM Studio desktop; Open WebUI web; HammerAI web, desktop and iOS.
One of five lists a phone app, and that row reorders the shortlist for anyone whose companion lives in a pocket. SillyTavern and Open WebUI are reachable from a phone browser, which is not the same as an app, and both runtimes are desktop by design. HammerAI's note adds that its desktop and mobile apps run models locally via Ollama.
The rule is blunt. One machine and a browser: four qualify, and the choice is about memory and interface. A native phone app as a hard requirement: the local shortlist is one product long, and the honest alternative is a hosted service, which this page does not cover.
Eleven of 154 is the number that should frame this
The directory holds 243 listings: 154 SFW and 89 adult-tier. Within the SFW set, 145 have a free tier, 110 record memory, 25 record voice calls, 25 record no advertising, and 11 can run locally.
Read those counts together and the shape appears. Free is the default, not a differentiator: 145 of 154, so free cannot separate one product from another. Memory is common but not universal at 110. Voice calls are scarce at 25. Local running is rarest at 11 — under one in fourteen of the SFW listings, the correction to carry into any article that calls self-hosting the way people use these products.
Two consequences. The eleven are worth browsing as a group: they are collected under the self-hosted list and the self-hosted tag, and this page covers five of them. And rarity is the argument. A feature 145 of 154 already have does not change a decision; a feature 11 of 154 have, and that decides whether a conversation ever leaves your machine, does. The same reasoning puts the local model tag ahead of any general best-of list for readers whose first question is where the data goes.
Decision rules
- Zero subscription, and you supply the model. SillyTavern, $0 monthly and $0 annual equivalent, verified 2026-09-14, with a runtime beneath it; the only one here listing voice calls.
- A GUI over open models on one machine. LM Studio at $0, verified 2026-09-30, free for personal use; its note separates team licensing for work.
- A local model server several apps share. Ollama at $0, MIT, verified 2026-09-30 — no memory, no voice, no group chat; it needs a frontend above it.
- The browser frontend with the most rows filled. Open WebUI at $0, verified 2026-09-29: memory, voice, images, custom characters, group chat.
- A free cloud allowance, a local mode and a phone app in one product. HammerAI: 600 cloud messages a day free, unlimited chat on locally run models, Pro at $9 a month with a $9 annual equivalent, verified 2026-09-29 — $108 a year, no annual saving to weigh.
- A native iOS app is non-negotiable. HammerAI is the only one of the five listing iOS; if it must also run fully local, confirm that path first.
- Deploying at work. The licence decides: AGPL-3.0 on SillyTavern, MIT on Ollama, team licensing on LM Studio, an enterprise tier on Open WebUI. Personal-use terms do not transfer to an office.
- Privacy is the actual reason. A local model is the requirement, not a self-hosted interface; any of these frontends pointed at a cloud API keeps the shell private and sends the conversation out.
- Price decides. This shortlist cannot answer it — the missing line item is your model spend. Estimate your token volume first.
Disclosure: our team also builds SoulChat, a hosted AI companion, which competes with the cloud half of this category. It is not part of this shortlist, and no ranking here depends on it. Prices, free tiers, platforms and verification dates above are recorded directory values; where a vendor publishes no figure, this page leaves it blank rather than estimating.