Autocomplete, chat, inline edit and code review in VS Code, on models you run. Any OpenAI-compatible server, on a laptop, in your data centre, or pooled from the team's own computers. No account, nothing phones home.
Free for individuals and teams of five. Team seats from $15 a month. Pricing.
250,000+ installs, Marketplace and Open VSX 3,600+ stars on GitHub in development since 2023
Where your code goes, autocomplete, inline edit, chat, and the gateway for teams. It has music; the sound is off until you turn it on.
Small, explicit features in the editor. Every one is a plain VS Code command you can rebind.
Shift+Tab and the chat model works in the workspace: reads, searches, asks the language servers, edits files and runs commands, each step shown in the reply. Commands ask before they run. Undo any file a reply changed. It follows AGENTS.md and takes MCP servers. Experimental; it wants a capable coder model.
docs →Fill-in-the-middle suggestions as you type, streamed as ghost text and cut off at a sensible end. Tuned to work well with a small model.
docs →Type @ to attach a file, a symbol, the problems panel, the git diff, the terminal, or a search of the workspace index.
docs →Ctrl+I, describe the change, review it as a diff in the editor. Accept or reject per hunk. Nothing is kept until you say so.
docs →A hybrid keyword and vector index of the workspace, reranked before it reaches the prompt and updated on save.
docs →Review the working tree, a branch against its base, or a GitHub pull request. The review lands in chat so you can ask follow-ups.
docs →Write a command from a description, shown before it runs. When one fails, twinny finds the file and line and offers the fix as a diff.
docs →There is nothing to sign in to because there is nothing on our side. Quick start.
From the VS Code marketplace. No sign-up, no key, no prompt to send telemetry.
$ code --install-extension rjmacarthy.twinnyOn first run twinny finds inference servers on localhost and offers them. Or paste any OpenAI-compatible endpoint.
provider http://localhost:8080/v1 ✓ connectedSuggestions appear as you pause. Open the sidebar to chat, press Ctrl+I to edit in place.
</> twinny · readytwinny-server sits between developers and your inference servers, or the team's own computers. Keys, usage, policy, plugins and an admin page, in one process you run.
# on the machine with the models $ npx twinny-server quickstart wrote twinny.gateway.json admin key (shown once): tsk_5f3a…9c listening on 127.0.0.1:8765 · 3 aliases · 1 key backend gpu-1 ok · gpu-2 ok · team 3 sharing admin: http://127.0.0.1:8765/admin # or the container: ghcr.io/twinnydotdev/twinny-server # then People → invite: one link per developer, opens VS Code
A developer flips Share this computer in VS Code. Their extension connects out to the gateway and serves the models on their local server to the rest of the team, least loaded first. No port to open, no firewall rule, no extra seat.
Switch a plugin on from the admin page and the gateway does more than serve completions. Watch your repositories and every open pull request is listed with its checks, approvals and where you stand. review now sends the diff to one of your own chat models; the review can be posted back to the host as a comment, a change request or an approval. Idle GPUs review new pulls and triage issues in the background, never slowing a developer down.
Requested from VS Code, approved with a short code, stored as a hash, revoked live. Offboarding is one click.
Requests, failures and tokens by developer, model and the machine that served them. Never the content.
Set the models once. Restrict which providers developers may add, keep a workspace on local backends, put a team system prompt before every chat. Decide what the agent may do: ask before every command, wait for review, keep out of paths.
Swap the model on the admin page. The next request uses it.
Only a named key can share. Sharers see one line of consent, admins see who serves what and can disconnect a machine.
Keep prompts and replies on your gateway for audit, review and training data. Off by default, disclosed to every developer.
Every admin change in a hash-chained log with filters and export. Read-only admin keys for the auditor. Prometheus at /metrics.
Give an alias a price per million tokens and usage shows what each developer, model and period cost.
When every slot is busy, requests wait briefly, oldest first. A team sharing one GPU sees a pause, not a failure.
Banks, exchanges, defence, healthcare, anyone shipping under NDA. The answers your security review will ask for, in one place.
| hosted | twinny | |
|---|---|---|
| Where prompts go | A vendor’s cloud | Your own servers |
| Whose GPUs | The vendor’s | Yours, or pooled from the team |
| Who picks the model | The vendor | You |
| Telemetry | On by default | None. There is no endpoint. |
| Air-gapped | No | Yes |
| Per seat, per month | Typically $19–39 | $0 to $25 |
Every claim above is checkable in the source. Every connection, what is kept and what an admin controls: the security page.
Every commit, issue and release is on GitHub, the licence check included. No sales team and no investors. The form below goes straight to the people who make twinny.
Every commit is public. The security review can read the source, the licence check and the audit log format before anyone signs anything. source · changelog · discussions · @twinnydotdev
A seat is an active key on your gateway. Revoke a key and the seat is free again. Everything in the extension and the gateway is in every tier; you pay for seats, policy, recording, plugins and support. Where the usual tools ask $19 a seat, twinny asks $15.
Pay by card. The licence token appears on your licence page and by email; paste it into the admin page. Renewals reissue the same licence. Prices in USD, issued by rjmacarthy.xyz. How licensing works.
A token by email: 25 seats, policy, recording and plugins, for 30 days. Paste it into the admin page. When it ends the gateway goes back to the free plan; nothing is deleted, and seats bought later carry on from there. One trial per address. How licensing works.
50 seats or more, an invoice or a purchase order, a named contact. Say who you are and how many developers; you hear back within one business day. Rather talk? Tick the box and say when suits.