SillyTavern alternatives: BYOK clients, hosted platforms and what migrates

The best SillyTavern alternative depends on which half of SillyTavern you are actually replacing: the interface, or the model supply. RisuAI is the closest like-for-like swap and the only actively-developed open-source client here that ships its own Android build. TavernAI 2 is the most aggressive migration path but is a closed-source binary with no published licence. Janitor AI, Chub AI, Character.AI and Backyard AI replace the model supply too, and the last two remove the bring-your-own-key field entirely.

SillyTavern itself is not going anywhere: the repository is AGPL-3.0, not archived, at 33,523 stars and 6,316 forks, and release 1.19.0 was published on September 14, 2026 (GitHub API, checked September 19, 2026). It has no API and no hosted service, so "alternatives" means one of two different products: a different bring-your-own-key front end, or a platform that sells you inference as well as an interface.

Start with why you are leaving

The phone row is the one worth spelling out, because SillyTavern ships no mobile app. Its own Android (Termux) install page routes you through Termux: install Termux from F-Droid rather than the Play Store, run termux-change-repo and pick a mirror group, pkg install git nodejs-lts, clone the release branch and run start.sh — a Node server on the handset. The same page notes Termux installs are not officially supported by the SillyTavern team.

What each alternative actually is, as of today

Repository facts below were read from the GitHub API on September 19, 2026. Two of these rows contradict what the roundups say, and both matter before you commit.

TavernAI is not the open-source option any more. SillyTavern was forked from a 2023 project of that name, but the repository at that path today was created on June 16, 2026 and its full contents are .github, .gitignore, CHANGELOG.md, CONTRIBUTING.md, README.md, SECURITY.md, assets, docs-site and locales — no source, no LICENSE, and the API reports licence null. The downloads are free to fetch, but neither the site nor the README states a price or licence terms, so treat the cost as unstated rather than as free software. What became of the original 2023 repository is something this page could not establish.

Backyard AI is not a local app. Faraday.dev announced on May 13, 2024 that it would transition all products and services to Backyard AI on May 20, 2024, and the thing it was known for is gone: desktop.backyard.ai states the desktop app is deprecated and no longer supported and points users to the hosted service. If you were considering it as a local-inference replacement, that product no longer exists.

Agnai needs a date, not an adjective. There is no archive notice and no deprecation statement, but the newest commit on the dev branch is dated June 13, 2026 and the repository has not been pushed since June 15, 2026 — roughly three months of silence when checked. That is cadence evidence; draw your own conclusion rather than accepting "abandoned" or "actively maintained" from anyone, this page included.

The bring-your-own-key field is not the same field twice

This is the part that breaks migrations in the first ten minutes. Every one of these apps has a box you paste a URL into, and they do not want the same string.

Two RisuAI details are worth knowing before you blame the endpoint. Its Format dropdown is an explicit enum — OpenAICompatible, Anthropic and OpenAIResponseAPI among others — so picking the wrong one produces a protocol mismatch rather than a helpful error. And a custom model's capabilities are not discovered: vision, prefill, streaming and prompt caching are hand-ticked flags, so they silently do nothing until you set them. It is also worth knowing where the "Autofill Request URL" rewrite does and does not apply: in the request builder the append is guarded by aiModel === 'reverse_proxy' && db.autofillRequestUrl, so a Custom Models entry never reaches it and its URL is passed through untouched. None of this was runtime-tested here; it is read from the main branch of kwaroran/Risuai on September 19, 2026.

Subscriptions versus your own token bill

Prices below were read from each vendor's own page on September 19, 2026. Where a figure is not published on a page this check could read, the cell says so rather than borrowing a number from a review blog.

Note what a $35 roleplay subscription actually reaches in late 2026: a line-up of open-weight models, several of them years old. That is the trade, and it is the reason the bring-your-own-key route survives. Three widely-quoted numbers are deliberately missing above — Chub's Mercury price and its supposed four-month billing blocks, and Janitor's free message cap — because they appear only on affiliate blogs and contradict or exceed what the vendors publish. Confirm any of them in that platform's own checkout.

What the same money buys as tokens

This is illustrative token arithmetic, not a measured cost and not a bill ceiling. Assume one generation sends a 12,000-token prompt — character card, system prompt and as much history as the context holds, all re-sent every turn — and receives a 250-token reply, with no re-rolls and no prompt caching. Rates are live Kunavo catalog prices per million tokens. The last column divides the cheapest subscription verified here, janitor+ at $12.99, by the estimated cost of one turn.

The spread is the point, not any single row: the same assumed turn costs about 98 times more on Claude Fable 5.1 than on GPT-5.6 Luna, so "is BYOK cheaper than a subscription" has no answer until you name the model. This is also not an equivalence: a subscription buys platform features, hosted character hubs and a fixed monthly number, and metered tokens buy none of those. Scale the arithmetic by your own turns per day before treating it as a budget, and note that context size drives it far harder than the headline per-million rate — double the history and you roughly double the input side.

Kunavo's catalog amount is a billing floor rather than a cap: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost multiplied by the applicable markup. Cache charges and external tools sit outside this example. There is no free tier and the minimum top-up is $10 in prepaid credit, which is a funding minimum, not a task fee or a subscription — see billing details. If a long chat is what you are budgeting for, prompt caching changes the input side more than any model swap does.

What migrates, and what you will rebuild

Cards travel because the format is shared. Character Card V2 was marked "Spec approved" on May 17, 2023, and names agnai, SillyTavern and RisuAI as the V1-compatible frontends — plus the ZoltanAI card editor — that it was aimed at; it does not certify that any of them implement it. Character Card V3 describes its card object as a superset of the V2 object "except the fields that are changed or removed", and adds a ccv3 PNG text chunk and a .charx zip archive. Everything wrapped around the card is where the work is.

The practical route into TavernAI 2 is its Mass Import, which you point at a SillyTavern root folder or a single SillyTavern user folder. That works because SillyTavern stores everything on disk: its administration docs describe a per-user folder at data/[user-handle] holding character files, conversation history and settings. Its own documentation also notes that only the default data layout is described there, so a relocated data root needs checking before you point anything at it.

Back up before you import, and keep the old install. Copy the whole data/ directory somewhere outside the SillyTavern folder first. An importer reads that folder, and a bad run is far cheaper to recover from a copy than from a card-by-card re-export. Do not uninstall SillyTavern until you have run a week of real chats in the new app — rollback is then just launching the old one, because its data never moved. That is also the honest reason to move cards by export rather than by deleting anything: a PNG you exported still opens in both.

Content rules and model availability come before price

Two things get conflated in every roundup. Platform policy varies: Backyard AI advertises fictional text and voice chats with no filters, and Chub AI's documentation describes its Mars models as uncensored, while Character.AI barred users under 18 from creating or talking to chatbots starting November 25, 2025, and Janitor AI announced on June 15, 2026 that users in Brazil, Australia and the UK will need to verify their age, calling it a legal requirement and naming Australia's Age-Restricted Material Codes under the Online Safety Act 2021 (in force March 9, 2026) and Brazil's Lei No. 15.211/2025 (in force March 17, 2026); the methods offered are facial age estimation, an ID scan or a k-ID AgeKey, and the post says there is no specific date yet. Do not read any of this as a uniform property of the category.

Model policy is separate and follows the model, not the app. Anthropic's Usage Policy, effective September 15, 2025, prohibits generating sexually explicit content and lists "Engage in erotic chats" explicitly. Putting a gateway URL in a different app's box does not change that. For unrestricted adult roleplay the honest answers are local inference, open-weight hosts, or a platform selling its own models — not a frontier key behind a new base URL. OpenAI's current adult-content wording could not be read for this page, so it is not characterised here.

Picking one, and wiring it up

If the interface is the problem, try RisuAI first and keep SillyTavern installed. If installation is the problem and you accept an unlicensed binary, TavernAI 2 has the only importer that reads your SillyTavern folder wholesale. If the bill is the problem, KoboldCpp with KoboldAI Lite or AI Horde takes it to zero. If you want to stop thinking about keys, a hosted platform is the answer and the custom endpoint stops being relevant. And if what you actually want is a general assistant, LibreChat and Open WebUI are the comparison — note that LibreChat is MIT while Open WebUI's licence forbids removing its branding above fifty end users in a rolling thirty days, and that LobeChat now ships as LobeHub.

For any of the bring-your-own-key routes, Kunavo is one OpenAI-compatible endpoint and one prepaid balance behind whichever front end you land on. These clients have not been runtime-tested against Kunavo's endpoint; a published setup guide is a configuration reference, not a compatibility test. Start from the SillyTavern setup for the base-URL form or the Janitor AI setup for the full-path form, then create a Kunavo account when you are ready to fund a key. Best API for SillyTavern covers model choice and cost per turn in depth if you are keeping the client and only changing the supply, and the OpenAI-compatible API reference explains what the endpoint expects.

FAQ

Repository facts, licences, release tags and commit dates were re-read from the GitHub API on September 19, 2026, and the Backyard AI, Character.AI, Chub AI, Janitor AI, SillyTavern and Anthropic pages linked above were fetched the same day. RisuAI, Agnai and TavernAI 2 behaviour is read from published source, locale files and documentation — nothing on this page was runtime-tested, and no client was connected to Kunavo's endpoint for this write-up. Kunavo token rates come from the live catalog; every dollar example here is illustrative token arithmetic.