ChatWise vs Cherry Studio: licence, agent gates and the shared API bill

ChatWise and Cherry Studio are both bring-your-own-key desktop clients — neither vendor sells you model access — so the only vendor fee that separates the two desktop editions is the client licence: ChatWise Pro is a one-time payment, listed at $29 against a struck-through $49, while Cherry Studio's Community Edition is $0 under AGPL-3.0. The choice that actually costs you something is the agent runtime, and the two are inverted: Cherry Studio's agent runs MCP servers but is documented as needing an Anthropic-protocol /v1/messages endpoint on its default runtime, while ChatWise's agent mode works over your provider but switches MCP off. Prices read from ChatWise's pricing page and cherryai.com on September 21, 2026.

Identity first, because both names collide badly. ChatWise here is the desktop AI client at chatwise.app — sold commercially, with no public source repository, credited on its own FAQ to a single open-source developer, and whose older v0 builds were published in the egoist/chatwise-releases repository. It is not chatwise.ai, a dating-reply generator, and not "Chatwise: AI Assistant" on the Apple App Store, which belongs to a different company. Cherry Studio here is the AGPL-3.0 desktop client from CherryHQ at github.com/CherryHQ/cherry-studio, whose site is cherryai.com. It is not cherrystudio.org, a product-photography business, and not cherrystudiocn.com, an unofficial lookalike that publishes a pricing page of its own with figures traceable to no Cherry source. Note also that the older cherry-ai.com domain now redirects, and the canonical documentation host is docs.cherryai.com.cn, with an English mirror.

Cherry Studio vs ChatWise: who should pick which

Workflow. ChatWise is the narrower, calmer product: a native chat client with assistants, web search, Artifacts, PDF handling and MCP tools, built for one person who wants it to open fast and stay out of the way. Cherry Studio is a workstation: many providers configured side by side, a knowledge base, an Agent workspace, and a Code CLI launcher that installs and points terminal coding agents at a provider you already configured, filtering out providers whose interface type a given CLI cannot use. If you want one app to hold every provider you own, that is Cherry Studio. If you want the thing you open to talk to a model, that is ChatWise.

Execution model and permissions. Read this before the price if you will let a model touch your machine. ChatWise's agent documentation gives the model five built-in tools — Bash, Write, Read, Edit and AskUserQuestion — loads Agent Skills from ~/.claude/skills and ~/.agents/skills, supports subagents and an off-by-default Auto Memory, and states plainly: "No sandboxing at the moment." It also disables MCP and Artifacts in the preview. Cherry Studio's Agent takes the opposite shape: MCP servers are bound per Agent, and every tool call runs under a permission mode. Its permissions page lists five — confirm each action, auto-accept edits, smart approval, plan-only and full access — and is explicit that the runtimes do not all offer the same set: only the Claude Agent runtime has all five, Pi has no plan-only mode, and DeepSeek Harness has no smart approval. An approval gate you can dial is a different risk posture from an unsandboxed shell tool, and it is the single biggest reason to choose between them.

Protocol and plan gates. Cherry Studio publishes a hard requirement that most comparisons miss. Its provider-type table classifies providers by protocol and marks the Anthropic-compatible type, /v1/messages, as the one Cherry Agent requires; its OpenRouter page spells out the consequence, advising readers not to run the Agent through OpenRouter because the Agent needs Anthropic's native protocol. So a gateway that only speaks chat completions can serve Cherry Studio's chat and knowledge base and still be unusable for its Agent. Read that requirement at the width the docs give it. It is stated for the Claude Agent runtime, which is the wizard's default; the documentation does not say which protocol the other two runtimes need, and describes DeepSeek Harness only as compatible with a variety of model services. Undocumented is not the same as unsupported, so treat the Anthropic requirement as settled for the default runtime and as an open question for Pi and DeepSeek Harness — and note that the model picker filters to the models a runtime accepts, which is where you will find out. ChatWise's gate is commercial rather than technical: the free tier is licensed for personal, non-commercial use — its FAQ gives use inside a corporation or for-profit organisation as the counter-example — so commercial use means buying Pro, and "Unlimited Tools (MCP)" sits on the Pro list, implying a free-tier cap whose number is published nowhere.

Cost path. Identical in shape and different in size. Neither vendor meters tokens, so both bills are: client licence, plus whatever your API provider charges, plus hardware if you run a local model. Cherry Studio's licence line is zero and it ships a free built-in provider, CherryAI, whose own documentation says the free model roster, quotas and rate limits are pushed from Cherry's servers, are adjusted from time to time, and that the list you see in the app is the only authority — so it is a nice way to start, not something to plan capacity around. ChatWise's licence line is a one-time payment, and its equivalent of a free lane is subscription reuse: its GitHub Copilot integration signs in by device code and draws on a Copilot subscription, and the homepage also advertises Claude Pro/Max reuse.

Migration cost. Asymmetric, and both directions have a trap. Leaving ChatWise costs you the licence: Pro is explicitly non-refundable, with the card-free 7-day trial given as the reason, so evaluate inside the trial rather than after paying. Leaving or upgrading inside Cherry Studio costs you data care: the V1-to-V2 differences page says migration is one-way, that the "Re-migrate" button deletes current V2 data before re-importing from the retained V1 data rather than merging, and that Anthropic OAuth credentials do not carry across, so those keys are re-entered by hand.

What each vendor charges

Read from chatwise.app/pricing, cherryai.com and Cherry's enterprise site on September 21, 2026. Because the $29 is a promotion with an end-date banner, budget against the $49 list price if you are planning more than a few days out, and confirm the amount in checkout.

Versions and platforms, and where the official sources disagree

Two version traps are worth naming because both will mislead a casual check. First, do not read ChatWise's version off its GitHub repository. egoist/chatwise-releases still shows v0.10.8 from March 19, 2026 and looks abandoned; it is not. The migration note explains that the desktop app was rewritten in Electron starting in 26.3.0, that this is where the YEAR.MONTH.BUILD_ID scheme comes from, and that the latest v26 build is downloaded from the ChatWise homepage; the homepage and changelog are where 26.9.0 appears. Second, Cherry Studio publishes its version in two places, so name the one you mean. The GitHub release page shows v2.1.0 published September 18, 2026 and cherryai.com/download showed the same v2.1.0 as the Stable Release when both were checked on September 21, 2026 — they agreed that day, but the download page labels one build "Stable Release" and files the rest under Previous Releases, so it is a curated channel and not the tag list — check both before you pin a version. The repository is not archived, and the GitHub API reported a push to it the same week. The ChatWise platform list above comes from the download links its homepage serves for 26.9.0, not from a published support matrix.

Bring your own key: what each endpoint box actually wants

Cherry Studio is Settings → Model Services → Add Provider, which opens an Add Custom Provider dialog. The important v2 change is that a provider is no longer one URL: the custom-provider page says the dialog offers OpenAI and Anthropic endpoints by default and adds OpenAI Responses, Gemini, an image-generation base URL and an image-editing base URL behind an extra-endpoint toggle it calls More Settings — the v2.1.0 client labels that toggle 更多端点 (More Endpoints), so trust the label in the app over either name. One provider entry can therefore carry both the chat-completions and the Messages endpoint of the same gateway, which matters given the Agent gate above.

The address field is a root, not an endpoint. The model-service settings page says that when a provider hands you https://xxx.com/v1/chat/completions you usually enter only https://xxx.com and Cherry Studio appends the rest, and that an address ending in # is used verbatim with no version path appended. This page deliberately publishes no literal Kunavo address for Cherry Studio. Kunavo's own Cherry Studio setup page and the client directory currently state that rule differently, and neither was settled by a live request here — so follow Cherry Studio's own append rule instead of either of them: enter the origin that your provider's /v1/chat/completions URL sits under, or paste that full URL and end it with # so nothing is appended. Then press Get Model List, add the models you want, and use Test to confirm the configuration before you rely on it. Multiple keys on one provider are supported and used in rotation from front to back, and a provider must be toggled enabled or its models never appear in the picker.

ChatWise is Settings → Providers → "+", then a preset or a custom configuration, then key and endpoint. Its custom-provider page states that it can integrate any OpenAI-compatible or Anthropic-compatible service; its changelog adds that 26.5.2 brought Responses API support to custom providers that are OpenAI-compatible. Models are fetched if the endpoint serves a /models route, or typed in by hand as a Model ID, display name, capability flags for image and video or audio input, a reasoning flag and a context length. That hand-declared metadata is the quiet hazard: nothing is discovered, so a wrong flag is a silently broken feature and any in-app spend figure reflects numbers you typed. The base-URL suffix for a ChatWise custom provider is not documented, and the only documented mapping — on its proxy page, for built-in providers — writes both the OpenAI and the Anthropic base with /v1, which is the opposite of the origin-only convention in Kunavo's Anthropic base URL reference. Read the field label in the app and confirm the first request lands rather than trusting either shape.

One ChatWise side effect is easy to trip over: its observability page supports exactly one provider, Cloudflare AI Gateway, and warns that the API Base URL for the respective provider must be empty or observability will not be enabled. Setting a custom base URL therefore turns ChatWise's only request logging off. If you need per-request visibility on a custom endpoint, get it from your provider's own usage ledger instead.

A worked estimate for the part of the bill both share

Because neither client meters tokens, the model bill is the same arithmetic on either side and the client choice does not change it. These are illustrative token estimates, not measured task costs and not a bill ceiling. Assume a short ask of 1,500 input and 700 output tokens, and a document turn of 100,000 input and 1,200 output tokens — the shape you get when a long PDF or a knowledge-base excerpt is pushed into context. Rates are live Kunavo catalog prices per million tokens; the monthly column assumes 200 asks and 20 document turns.

Two readings. First, the document turns dominate: under these assumptions the 20 long-context turns are about two-thirds of the month on every row, even though they are a tenth of the requests. The setting worth auditing in either client is therefore how much history and how many attachments go into each request — not which client you bought. Second, put the licence next to it. At $6.24 a month on Claude Sonnet 5, ChatWise's $29 one-time Pro payment is about 4.6 months of tokens, and it never recurs; on a cheaper lane such as Claude Haiku 4.5 at $1.25 a month, the licence is the larger number for a while. The client fee is a rounding error for a heavy user and the whole decision for a light one, so work out which you are before letting $0 decide it.

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. The minimum top-up is $10 in prepaid credit, which funds a balance rather than buying a plan — see billing details. Cheapest listed rate and cheapest way to finish a task are different questions; AI cost optimization covers the difference, and prompt caching covers the discount that long-context turns like the second column can attract.

Feature comparison

Compiled from ChatWise's MCP documentation, its documents page and agent page, and Cherry Studio's Agent tools page and create-agent page, all read on September 21, 2026. Cherry Studio's marketing copy and its documentation quote different provider counts and name model versions that have since moved on, so treat any count or model name on either vendor's homepage as stale until the docs confirm it.

If you are bringing a key to either one

Kunavo publishes a setup page for Cherry Studio and none for ChatWise, and neither client has been runtime tested against api.kunavo.com — a published configuration reference is not a compatibility test, and everything above was read from each vendor's own documentation rather than from a running install. What is structural rather than tested: Cherry Studio documents an Anthropic-protocol /v1/messages endpoint as what Cherry Agent needs on its default runtime, and Kunavo exposes such an endpoint, documented at the Messages API reference. That is a reason a gateway's protocol coverage matters to this comparison; it is not a claim that the pairing has been exercised end to end.

The sane order is: keep your current route working, add the new one beside it, run one bounded task, then read the charge your account actually recorded for it. Start at the Cherry Studio integration guide, confirm the address you enter follows Cherry Studio's documented append rule before you save it, and create a Kunavo account when you are ready to fund a key. For the wider client list and the endpoint shapes each one expects, see the AI client directory and the OpenAI-compatible API guide.

FAQ

Every source cited above was checked on September 21, 2026: ChatWise's homepage, pricing page and changelog, and its custom-provider, agent, documents, MCP tools, proxy, observability, GitHub Copilot and v26 migration documentation; Cherry Studio's homepage, download page, enterprise site, v2.1.0 GitHub release and GitHub API record; and Cherry Studio's provider-type table, custom-provider page, model-service settings page, OpenRouter page, CherryAI page, Code CLI page, V1-to-V2 page and its agent-workspace create-agent, tools and permissions pages. ChatWise's pricing FAQ, its changelog and Cherry Studio's download page all render with client-side script, so a plain fetch of any of them returns nothing useful; those three were read in a browser. The four same-name products named in the second paragraph were each opened on the same date to confirm they are different products; they are described rather than linked on purpose. Cherry Studio's Chinese documentation was read in the original where no English mirror exists, so the permission-mode and tool names here are translations of Chinese labels rather than strings you will see in an English build. Neither client was installed or pointed at Kunavo, and no claim here rests on a request either of them made. Kunavo token rates come from the live catalog, and every dollar example here is illustrative token arithmetic rather than a measured task cost.