Twinny vs Continue: what changed, and which one to use now

Twinny and Continue are not competing for the same job any more: Continue is finished software that still runs, and Twinny is a narrower tool that is still shipping. continue.dev's own page title reads "Continue (acquired by Cursor)" and its README calls the repository no longer actively maintained and read-only, while Twinny pushed versions 4.1.1, 4.1.2 and 4.1.3 on a single day, 2026-09-18. But Twinny deliberately has no agent mode and no MCP tools, so it replaces some of what Continue did and not all of it.

One disambiguation first, because the name is crowded. Two unrelated companies trade as Twinny: a Spanish enterprise-automation platform at twinny.es and a Korean robotics firm at twinny.ai, both live when checked on September 19, 2026. Neither one's prices or funding belong on this page. The product here is the VS Code extension at twinny.dev, published as rjmacarthy.twinny from the twinnydotdev/twinny repository, documented at twinny-docs.

What actually happened to Continue

The short version is that the product ended and the code did not. Continue's README states it plainly: the repository "is no longer actively maintained and is read-only for all users." The homepage's meta description is one sentence — Continue has been acquired by Cursor. Read-only does not mean archived, though, and this is the distinction most write-ups get wrong.

All rows checked September 19, 2026. Two of them deserve emphasis before you plan anything. First, the README says the final 2.0.0 covered the VS Code extension, the CLI and the JetBrains plugin, and the published artifacts disagree with it on two of the three: the CLI never went past 1.5.47 and the JetBrains plugin never past 1.0.67. This guide reports that discrepancy rather than explaining it, because no primary source explains it. Second, the documentation site outliving the hub is a live trap: docs.continue.dev still teaches the uses: <owner>/<slug> syntax for pulling a model or rule from a host that no longer resolves. A fully local config.yaml with explicit provider, apiBase and apiKey entries is now the only shape guaranteed to resolve.

Who should pick which

Pick Twinny if what you actually did in Continue was autocomplete, chat about your code, inline edits, review and commit messages — and you want a codebase someone is still touching. The repository is MIT, not archived, and its most recent commits on September 19, 2026 were the version bumps 4.1.1 through 4.1.3, all dated 2026-09-18. Pick it especially if you run models on your own hardware: Twinny's whole design centre is a local server, and its fill-in-the-middle autocomplete works well with a 1.5B to 7B base model.

Pick Continue, or rather keep it, if you depend on something Twinny does not have. Agent mode and MCP tool servers are the clearest case — Continue's MCP documentation says MCP "can only be used in the agent mode", so both stand or fall together. The CLI and the JetBrains plugin are the other case: Twinny requires Visual Studio Code 1.93 or newer, or a compatible build such as VSCodium, and ships no command-line client and no JetBrains edition. A frozen Continue install keeps working; it just will not improve.

The execution model is the real dividing line, and it is not a feature gap that a release could close. Twinny's FAQ answers "Does twinny run an agent that edits my files on its own?" with "No. Every change is something you asked for and can see before it is kept: a suggestion you accept, a diff you accept, a command you confirm. This is by design." If you were using Continue to hand off a multi-step task and review the result, Twinny is a downgrade you will feel on day one. If you were using Continue in chat and edit mode and found agent mode unnerving, Twinny is the same workflow without the part you did not want.

Migration cost is low but not zero. Continue puts every model in one YAML file with a roles list; Twinny has no config file at all and builds providers one job at a time in the sidebar, so a single Continue entry carrying three roles becomes up to three Twinny providers. Nothing in Twinny reads a Continue config. Budget the rewrite at about an hour, plus rebuilding the workspace index — vectors from different embedding models do not mix, so changing the embedding model means reindexing.

On cost, both clients are free and neither is the bill. Continue has no price at all: continue.dev carries no pricing page any more, and the only thing its FAQ says about the software is that the Apache-2.0 source and documentation remain on GitHub. Twinny's extension is MIT and free, but its team gateway is where a number appears, and the two halves are easy to conflate.

Feature and licence, side by side

Read from the two Marketplace listings, the repositories and the projects' documentation on September 19, 2026. Install counts drift daily and are a record of history, not of current use — Continue accumulated its number over three years as a maintained product. Note also that Twinny's Marketplace one-liner ("Locally hosted AI code completion plugin for vscode") and its GitHub description are both older than the positioning on twinny.dev, which now leads with the team gateway. Older articles describing Twinny's peer-to-peer "Symmetry" network describe a removed feature; its FAQ says Symmetry has been removed and replaced by Devices, which links your own machines only.

What each one costs

Twinny plan prices read from its licensing and seats page on September 19, 2026; Continue's position from continue.dev itself. Two honest caveats. A seat is an active access key rather than a person, and revoking a key frees the seat immediately. And these are published prices, not prices this guide put through checkout — the site names rjmacarthy.xyz as the licence issuer, and the seller entity behind that was not investigated. Twinny's own homepage comparison also quotes a "typically $19–39" per-seat figure for hosted assistants; that is its marketing, not a figure verified here.

What either client can reach on an OpenAI-compatible gateway

This is where the two tools stop being interchangeable in a way no feature table shows. Both split work into jobs, and only some jobs can talk to a chat-completions gateway at all.

Chat fits on paper. Twinny's providers overview states that for chat the API path is the base and Twinny appends /chat/completions, so an API path of /v1 against api.kunavo.com lines up with Kunavo's own route. The catch is which preset: Twinny's hosted APIs page says chat with a hosted API "goes through the vendor's SDK to its fixed endpoint, so the hostname, port and path fields are hidden." The OpenAI and Anthropic presets still ask for a key — they send it to that vendor, not to an address you pick, which is why a gateway key does not belong in them. The address fields stay visible on the local-server presets, and the generic "OpenAI-compatible server" preset is the one meant for a server twinny ships no preset for. Paste https://api.kunavo.com/v1 into Hostname and the form splits it into protocol, host, port and path for you; the key is sent as an Authorization: Bearer header.

Autocomplete does not fit, for two independent reasons. Twinny's generic preset drives completion through a /v1/completions-style route, and Kunavo has no such route — there is no legacy text-completions endpoint in its API, only /v1/chat/completions and the other chat-shaped surfaces. Separately, Twinny's supported models page is explicit that only models trained with fill-in-the-middle tokens can complete between what sits before and after the cursor, and that instruct models "tend to chatter or explain instead of completing." Kunavo's catalog is instruct and chat models; none of them is a fill-in-the-middle base model. Either blocker alone settles it. Continue's own autocomplete guidance points the same way without using Twinny's vocabulary: for that role it recommends Codestral and Qwen2.5-Coder at 1.5B and 7B, and warns that thinking-type models are not suited to it. Serve autocomplete from a small local model — that is the configuration both projects recommend anyway.

Embeddings are not served by Kunavo at all. No model in the catalog carries the embedding endpoint, so /v1/embeddings resolves a model and rejects the request. Point Twinny's embeddings provider, or Continue's embed role, at a local server or an outside embedding provider. Twinny's docs recommend nomic-embed-text for this.

name: My Config
version: 0.0.1
schema: v1
models:
  - name: Claude Sonnet 4.6
    provider: openai          # the protocol, not the vendor
    model: claude-sonnet-4-6
    apiBase: https://api.kunavo.com/v1
    apiKey: <your Kunavo key>
    roles: [chat, edit, apply, summarize]
    capabilities: [tool_use]  # only if agent mode stays unavailable

Two notes on that block. provider: openai names the wire protocol rather than the vendor, so it applies to any endpoint implementing /v1/chat/completions. And the capabilities line is the one non-obvious part: Continue's capabilities documentation says it detects tool_use from the provider and model name, and that "you cannot override autodetection — you can only add capabilities." A gateway serving an id Continue's table does not recognise gets no tool_use, and the capabilities page lists agent mode being unavailable as exactly what that causes; adding the line switches it on, and nothing switches it back off. Note also that Kunavo's own Continue setup page ships a second entry using a chat model in the autocomplete role. That configuration has not been runtime-tested by anyone on record, and the models Continue's own autocomplete guidance names for that role are code-completion models rather than chat ones — treat it with the caution this section describes rather than as a verified recipe.

A worked token estimate for the chat-shaped work

Because autocomplete cannot reach the gateway, the token bill for either client against Kunavo is chat traffic only: chat answers, inline edits, code review and commit messages. That makes the arithmetic smaller than a completion-heavy assistant's, and it is worth seeing the size before choosing a model. Assume a working day of 12 chat or edit turns averaging 18,000 uncached input tokens and 1,200 output tokens each — 216,000 input and 14,400 output in total. Those assumptions are illustrative; substitute your own before budgeting. Rates are live Kunavo catalog prices per million tokens.

These are illustrative token arithmetic, not measured task costs and not a bill ceiling. The cheapest listed rate and the lowest cost to finish a task are different claims: a model that needs a second attempt at your review can cost more than a pricier one that lands first time. 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 times the applicable markup. Cache charges and anything you run locally sit outside this example. The minimum top-up is $10 in prepaid credit, a funding minimum rather than a task fee or a subscription; see billing details.

Setting it up

Kunavo publishes a setup guide for Continue but none for Twinny, and a published configuration reference is not a compatibility test — neither client has been runtime-tested against this endpoint. Keep a working route open while you try one, run a single bounded task, then read the charge your account recorded for it. For Continue, start at the Continue integration guide and read it alongside the caveats above. For Twinny, use the generic OpenAI-compatible provider described here and Twinny's own Test provider button, which sends a small request for that provider's job and shows either success or the server's error along with the URL it called — that is the fastest way to tell a wrong path from a wrong key. Create a Kunavo account when you are ready to fund a key.

Comparing further afield? OpenAI-compatible API covers what the protocol does and does not carry, and Cline vs Claude Code compares two agents if Continue's agent mode is the part you cannot give up.

FAQ

Checked on September 19, 2026 by fetching them directly: both GitHub repository APIs and commit lists, the two VS Code Marketplace listings, the npm entries for @continuedev/cli and twinny-server, the JetBrains plugin listing, continue.dev and its README, a DNS lookup for hub.continue.dev, docs.continue.dev's OpenAI-provider, capabilities and MCP pages, and Twinny's introduction, providers, hosted-APIs, supported-models, licensing and FAQ pages. Not checked: any runtime request from either client to Kunavo, and Twinny's checkout. Kunavo token rates come from the live catalog; every dollar example here is illustrative token arithmetic.