Every release, newest first.
The extension and twinny-server are built from one tree and share a version number. This page is the repository's CHANGELOG.md, one page per release, and an RSS feed.
- 4.3.15 2026-10-16
Extension release: a reworked sidebar, agent mode that can be steered while it works, and better handling of small local models.
- 4.3.14 2026-10-15
Extension release: a DemonRoute preset among the hosted chat providers, contributed by @haterfactory (#545). It fills in api.demonroute.com over https with /v1 and shoukewei/Qwen3.8-27B-OBLITERATED22-AGENT as the model; any model on your key can be picked instead. Chat only, no autocomplete or embeddings. Its description is translated into every language the sidebar ships.
- 4.3.13 2026-10-15
twinny-server release: the admin page's Buy seats link goes to twinny.dev's pricing instead of a Stripe link built into the gateway, so a change of price or checkout needs no new release. Team seats are now $15 a seat a month, billed yearly.
- 4.3.12 2026-10-15
Extension release: a ChatGPT Plan provider, contributed by @CoderLambert (#540). twinny-server refuses the new provider kind, which needs the editor, and masks more secrets in its logs.
- 4.3.11 2026-10-15
Extension and twinny-server release: completions work against llama-server in router mode.
- 4.3.10 2026-10-15
Extension release: @ mentions in the chat composer list matches as you type, and @symbol mentions attach the symbol. twinny-server carries the version number only.
- 4.3.9 2026-10-14
Version number only; no changes from 4.3.8.
- 4.3.8 2026-10-14
Extension and twinny-server release: agent mode's changes can be undone from the reply, it follows the project's AGENTS.md, it sees the errors its edits cause, and a team's admin can set what it may do. Also a round of fixes from an audit of agent mode, two of them security fixes to the git tool.
- 4.3.7 2026-10-07
twinny-server release: access tokens for the GitHub, GitLab, Gitea and Bitbucket plugins are saved once and shared, so an expired token is replaced once. The extension carries the version number only.
- 4.3.6 2026-10-06
Extension release: the chat composer suggests how to finish what you are typing. twinny-server release: sign-in's per-address limit works behind a reverse proxy.
- 4.3.5 2026-10-02
Extension release: settings, labels and a command that did not do what they said. twinny-server carries the version number only.
- 4.3.4 2026-10-02
Extension release: the chat's provider and model dropdowns open again. twinny-server carries the version number only.
- 4.3.3 2026-10-02
twinny-server release: Recordings folds an agent conversation's steps into one row, and the tests no longer touch the real data directory. The extension carries the version number only.
- 4.3.2 2026-10-02
Extension release: agent mode's steps show inside the reply, commands run in the background with their output in the chat, and commands can be approved from the keyboard. twinny-server carries the version number only.
- 4.3.1 2026-10-01
Extension release: code in agent mode's steps wraps properly. twinny-server carries the version number only.
- 4.3.0 2026-10-01
Extension release: agent mode switched from the chat, messages queued while a reply runs, terminal-style chat keys, and a fix for Anthropic tool calls. twinny-server carries the version number only.
- 4.2.10 2026-09-30
Extension release: twinny starts in WSL again, and the Embeddings tab holds still while indexing.
- 4.2.9 2026-09-29
twinny-server release: twinny-server --version and the gateway's /twinny/v1 responses report the right version. The 4.2.8 package was published with a bundle built before the version bump, so it called itself 4.2.7. Publishing now refuses a cli.js that was not built for the package's version.
- 4.2.8 2026-09-28
Extension and twinny-server release: plugins shared with developers, opened from VS Code already signed in, an approve button on the pull page, and Qwen3-Coder completions through a gateway.
- 4.2.7 2026-09-24
Extension and twinny-server release: the completion model is loaded before you type, and the gateway shields secrets for every client.
- 4.2.6 2026-09-24
Extension release: a secret shield keeps credentials out of prompts that leave the machine. twinny-server carries the version number only.
- 4.2.5 2026-09-23
Gateway release: pull-request reviews stop ending mid-sentence, say when they were cut, and can be asked about. The extension carries the version number only.
- 4.2.4 2026-09-23
Chat release: replies carry their details, and the chat gains the controls it was missing. twinny-server carries the version number only; the gateway is unchanged.
- 4.2.3 2026-09-23
Small release: code completion works with Qwen3-Coder. twinny-server carries the version number only; the gateway is unchanged.
- 4.2.2 2026-09-23
Small release: prompt templates (~/.twinny/templates) are sturdier. twinny-server carries the version number only; the gateway is unchanged.
- 4.2.1 2026-09-22
Small release: the extension tells the person who could run a gateway that one exists, and the listings say what the product costs.
- 4.2.0 2026-09-21
Everything in it is on the gateway side; the extension only learns to send the open workspace's name (for routing rules) and to search the shared context index.
- 4.1.3 2026-09-18
- Gateway discovery now names the backend model behind each alias, and the extension picks the fill-in-the-middle prompt format from it instead of from the alias name. An alias called coder in front of Codestral gets the Codestral format.
- 4.1.2 2026-09-18
- Mistral: the completion endpoint gets the model in the request body, and text-only chat messages go as a plain string rather than a parts list, which Mistral refused. Images keep their parts.
- 4.1.1 2026-09-18
Teams. One gateway on your own hardware serves twinny to the whole team.
- 4.0.20 2026-09-17
- Chat: Add file to context did nothing (#503).
- 4.0.19 2026-09-14
- Chat: code blocks render and copy better.
- 4.0.14 to 4.0.18 2026-09-13
- Workspace index: models that truncate at 256 tokens are given embedding windows instead of whole chunks; identifiers are split for the keyword column; follow-up questions search with their question; open files are favoured; hits are widened to neighbours. Workspace search sources show under chat replies. - Autocomplete gets the definitions and signatures of what is being typed from the language server. - Providers: validation, a Test provider probe on every card, presets, import and export; the provider layer is laid out by feature. - Logging through a VS Code LogOutputChannel.
- 4.0.10 to 4.0.13 2026-09-10
- Mentions: @git, @terminal and @symbol join @files, @problems and @workspace. - Terminal: fix the last error with the output attached, and write a command from a description, shown before it runs. Needs VS Code 1.93 shell integration. - Workspace index rewritten: an incremental index with a manifest, hybrid BM25 and vector search, a reranker in worker threads, @workspace or automatic mode. - Dependencies updated.
- 4.0.7 to 4.0.9 2026-09-07
- Inline edit: Ctrl+I streams a merged diff into the editor with Accept and Reject per hunk (CodeLens and keys); Fix with Twinny quick fix. - Autocomplete takes context from the language server and from recent edits (diff hunks against a baseline per document). - Tests written next to the code they cover.
- 4.0.0 to 4.0.6 2026-09-07
The 4.x rewrite.