OmniRoute by Cheaper Inference for GitHub Copilot Chat
Created by Diego Souza
1200+ AI models in your Copilot Chat — free & forever free.
340+ providers, 90+ with free tiers, one endpoint. MIT open source.
🧩 Install from the VS Code Marketplace
🔓 Install from Open VSX (Cursor, Windsurf, VSCodium, Theia, code-server…)
🌐 OmniRoute on GitHub •
🔌 Extension source •
🐛 Issues
Don't replace Copilot — power it up. No new sidebar to learn, no new chat UI. This extension drops every model from your OmniRoute server — Kimi, Claude, GPT, Gemini, GLM, DeepSeek, Qwen, Llama and hundreds more across 340+ providers, 90+ of them with free tiers that stay free — straight into the model picker of the Copilot Chat you already use. Including VS Code Insiders.
🆓 No Copilot subscription required. Since VS Code 1.122, provider models work without a GitHub sign-in and without any Copilot plan. VS Code + OmniRoute + this extension = a fully working AI chat with agent mode, for free.
Why this extension?
- 1200+ models, one picker. OmniRoute unifies 340+ providers (OpenAI-compatible, Anthropic, Gemini, Ollama, local, OAuth-based free tiers…) behind a single endpoint — 90+ providers are free, and free forever. Every model it serves shows up in your Copilot Chat model dropdown.
- Agent mode, tool calling, MCP, instructions — all of it still works. This plugs into VS Code's native language-model provider API, so Copilot's entire stack now runs on the model you choose.
- Vision included. Models flagged as vision-capable in the OmniRoute catalog accept image attachments directly in chat.
- Reasoning you can read. Thinking models show their reasoning in the chat as it streams, above the answer.
- Prices in the picker. Each model shows what OmniRoute says it costs, its context window, and for a combo the strategy and the steps behind it.
- Combos & auto-fallback. OmniRoute combos (priority, round-robin, cost-optimized, fusion…) appear as regular models — pick one and get automatic failover across providers behind the scenes.
- Online at a glance. A status-bar dot and an Activity Bar panel show whether your OmniRoute server is reachable, how many models it serves, and where to configure everything — URL and API key included.
- Speaks your language. The UI ships in 42 languages, mirroring OmniRoute's own catalog — VS Code picks the one matching your display language automatically.
- Configure your other tools too. One command configures Codex CLI, Claude Code, Cline, Continue, Cursor, Aider and more to use OmniRoute — powered by the
omniroute CLI under the hood.
- Secure by default. The API key lives in VS Code's SecretStorage (OS keychain), never in
settings.json, and it is only sent to the server you saved it for. A repository you open cannot point the extension somewhere else.
Getting started
1. Run OmniRoute (60 seconds)
npm install -g omniroute
omniroute # dashboard at http://localhost:20128
Add your providers/keys in the dashboard — or use the built-in free ones. Full guide: github.com/cheaper-inference/OmniRoute.
Already running OmniRoute somewhere else (home server, VPS, tunnel)? Point the extension at it — see below.
2. Install this extension
Search for "OmniRoute" in the VS Code Extensions view — it resolves from the
Marketplace on VS Code
and from Open VSX on forks like Cursor,
Windsurf and VSCodium. Or grab the .vsix from the releases.
3. Pick a model
- Open Copilot Chat and click the model picker
- Choose Manage Models… → OmniRoute
- Tick the models you want — they now live in your picker
That's it. If OmniRoute runs on the default http://localhost:20128, there is nothing to configure.
The same four steps are in VS Code itself: open the Welcome page and pick Get started with OmniRoute (or run Welcome: Open Walkthrough…). The last step points chat titles, commit messages and inline chat at an OmniRoute model too (chat.utilityModel, chat.utilitySmallModel, inlineChat.defaultModel).
Remote server / API key
Click the OmniRoute icon in the Activity Bar (or run OmniRoute: Manage Connection, or click the status-bar dot) to open the connection panel:
- Server URL — e.g.
http://my-vps:20128 (the /v1 suffix is added automatically)
- API key — only if your OmniRoute requires one (
REQUIRE_API_KEY); stored in the OS keychain
- Save & Test — instant feedback with the live model count
Two rules keep the connection yours:
- The server URL comes from your user settings. In a remote window (WSL, SSH, a dev container) the remote machine's settings count too. A value in a workspace's
.vscode/settings.json is ignored, so opening a repository cannot point the extension at another server.
- The API key is tied to the server it was saved for. If the URL changes, the key stays put until you approve the new server, with Save & Test or Use the key here on the warning. A server reached over plain
http:// outside your own machine gets a one-time warning before the key travels.
One API key per workspace
Teams that give each project its own OmniRoute key can keep that key with the workspace. With a folder or workspace open, tick Use this key only in this workspace in the panel and save the key. It is stored in the OS keychain for that workspace and is used whenever that workspace is open; every other workspace keeps using your global key. The line under the field says which key is in use, and Clear key removes the one in use.
- A workspace key is tied to its server like the global one, and each key has its own approved servers.
- Moving or renaming the folder (or the workspace file) makes it a new workspace: the key stays stored but unused until you save it again there.
- Configure Coding CLI asks before it hands a workspace key to a tool, because the tool keeps one configuration for all your projects.
- To group usage by project on the server even when workspaces share a key, turn on
omnicopilot.sendProjectName: chat requests then carry the workspace name. It is off by default.
Troubleshooting the connection
When the server cannot be reached, the panel says why instead of only "Offline":
| The panel says |
Usual cause |
| Nothing is listening (connection refused) |
OmniRoute is not running, or the URL points at the wrong machine or port |
| No answer within a few seconds |
A firewall, or a host that is down |
| The network cannot reach the address |
No route from this machine: a VPN that is off, or another subnet |
| The host name could not be resolved |
A typo in the URL, or DNS |
| The request failed at a proxy |
VS Code sent the request through a proxy. Add the host to http.noProxy |
| The TLS certificate was rejected |
A self-signed or expired certificate on an https:// server |
| Answered HTTP n, which is not what an OmniRoute server answers |
Something else is listening at that address |
Two situations explain most reports:
- Remote windows. In WSL, over SSH, in a dev container or a Codespace, the extension runs on the remote machine.
localhost is that machine, not your own computer. Use an address the remote can reach, and set it in the panel from that window.
- Proxies. VS Code routes extension requests through its proxy settings, LAN addresses included. If
curl reaches the server from a terminal and the extension does not, add the host to http.noProxy, for example "http.noProxy": ["192.168.2.4"].
Run OmniRoute: Configure Coding CLI and pick a tool — the extension drives the omniroute CLI to generate ready-to-use profiles:
| Tool |
What you get |
| Codex CLI |
codex --profile glm52 style profiles in ~/.codex |
| Claude Code |
omniroute launch --profile <name> launch profiles |
| Cline / Roo / Kilo |
Extension settings pointed at OmniRoute |
| Continue / Cursor / Aider / Goose / Crush / OpenCode / Qwen Code / Gemini CLI / Zed / WhyCodes / 5dive |
Tool-native config |
Settings
| Setting |
Default |
Description |
omnicopilot.baseUrl |
http://localhost:20128 |
OmniRoute server root (local or remote) — /v1 is appended automatically. Read from your user settings; a workspace value is ignored |
omnicopilot.modelFilter |
(empty) |
Substring/regex to limit which models are listed |
omnicopilot.maxOutputTokens |
16384 |
Output budget reserved per response |
omnicopilot.defaultContextLength |
128000 |
Context assumed when the catalog omits it |
omnicopilot.statusBar |
true |
Show the connection dot |
omnicopilot.statusBarUsage |
true |
Show how much of the key's daily limit is left next to the dot |
omnicopilot.healthCheckIntervalSeconds |
30 |
Probe frequency |
omnicopilot.dashboardOpen |
external |
Open the dashboard in the browser, or in a VS Code tab (editor) — see below |
omnicopilot.defaultReasoningEffort |
(unset) |
Effort for models with extended thinking — none/low/medium/high/xhigh/max; see below |
omnicopilot.cliPath |
omniroute |
Path to the OmniRoute CLI (a machine setting; put a path with spaces in quotes) |
omnicopilot.exposeToAgentsWindow |
false |
Experimental: also list the models in the Copilot Agents window — see below |
omnicopilot.sendProjectName |
false |
Send the workspace name with chat requests so OmniRoute can group usage by project. Read from your user settings only |
Reasoning effort
Models with extended thinking (GPT-5.x, Claude with thinking, DeepSeek R1, Gemini
Thinking…) accept a reasoning tier.
On VS Code 1.120 and newer those models get a Thinking effort menu in the model picker,
with the tiers the model offers: Off, Low, Medium, High, Extra high and Max. Pick one and it is
sent to OmniRoute as reasoning_effort on every request to that model.
omnicopilot.defaultReasoningEffort sets what is used until you pick something, and covers
editors that show no menu. It uses OmniRoute's vocabulary (none, low, medium, high,
xhigh, max), and the server steps down a tier a model does not implement, so asking for
max is always safe.
The default is applied only to models the catalog marks as reasoning-capable: sending the
field to a model without thinking support is ignored at best and rejected with a 400 at worst.
A choice made in the picker always wins over the setting.
Agents window (experimental)
Turn on omnicopilot.exposeToAgentsWindow to also register every tool-calling OmniRoute model
for the Copilot Agents window (the Copilot CLI agent host). VS Code scopes these entries to
agent-host sessions, so they appear in the Agents window's model picker instead of the regular
chat picker — the regular picker keeps working unchanged.
Prerequisites, both on the VS Code side:
- Enable the experimental VS Code setting
chat.agentHost.byokModels.enabled.
- Restart the agent host (or reload the window) after flipping the settings.
This rides a proposed VS Code API (targetChatSessionType), which is why it ships opt-in
and default-off: the mechanism can change under us in a VS Code update. If the entries show up
duplicated in the regular picker instead of the Agents window, your VS Code build does not
support the proposal yet — turn the setting back off.
Your usage in the panel
When OmniRoute runs a recent enough version, the Activity Bar panel shows a My usage
section — your key's own daily/weekly spend against its limit, the reset time, and the quota
of each provider connection (Codex, Claude, OpenCode…) side by side.
It reads the self-service endpoint GET /api/usage/om-usage?format=json with your connection's
API key — the same data as the terminal @@om-usage command, never the management surface.
Each line shows what is left and when it resets. A quota that is a prepaid balance (DeepSeek,
Kilo, Lyceum…) shows the money left, such as $0.37 left, instead of a percentage.
When the usage command is not enabled for a key, the panel reads the key's own daily and weekly spend
from OmniRoute's self-service endpoint (GET /v1/me/status) instead, so most keys see their numbers
without anyone changing a flag. Provider quotas still need the usage command.
A key with a daily limit also shows how much of it is left in the status bar, such as
OmniRoute · 76%. It is read at most once every five minutes, and not at all while the status bar is
hidden; omnicopilot.statusBarUsage turns it off.
Three states are normal and distinct:
- "Usage is not enabled for this key" — the key lacks the
allowUsageCommand flag, which
is off by default, and the self-service endpoint is not available for it either. An admin
turns the flag on per key in OmniRoute's API-key manager. This is a setting, not an error.
- "nothing is cached yet" — the key is allowed but the server has no learned quota for the
connection yet. It fills in after the first requests.
- No section at all — no API key is in use, the usage call failed, or the OmniRoute server
predates
?format=json; the panel hides the section rather than guess.
Requires OmniRoute with the usage JSON endpoint (v3.8.50+). See
docs/CATALOG.md for the server version the feature matrix assumes.
Dashboard inside VS Code
Set omnicopilot.dashboardOpen to editor to open the OmniRoute dashboard in a VS Code tab
instead of your browser. This needs the server to allow embedding, which is opt-in via
DASHBOARD_ALLOW_EMBED=vscode (landed in
[OmniRoute #10273](https://github.com/cheaper-inference/OmniRoute/issues/10273)).
⚠️ That flag is read at build time — Next.js bakes the response headers into the route
manifest — so exporting it in front of an already-installed server does nothing. It works on a
build from source (DASHBOARD_ALLOW_EMBED=vscode npm run build), not on the prebuilt
npm install -g omniroute package or the official Docker image. Full matrix in the
OmniRoute guide.
Without an embed-enabled build the page refuses to frame; the extension detects that from the
response headers and falls back to the external browser, so nothing breaks either way.
How the model list is built
Curious why the picker shows the number of models it shows, or why a provider you never
configured is in there? → docs/CATALOG.md explains the duplicate-prefix
mode, the non-chat filter and the free/keyless providers, with measured numbers.
Good to know
- Chat, agent mode and utility tasks run through your OmniRoute models. Inline code completions and embeddings-based features are outside VS Code's provider API and still require GitHub Copilot.
- A cut-off answer fails loudly. When a model hits its output limit in the middle of a tool call, or spends all of it on reasoning, the request fails and says so instead of running the tool with empty input or returning nothing. Raise
omnicopilot.maxOutputTokens if it happens often.
- Temporary errors are retried. A rate limit or a busy provider is retried up to twice before you see an error, as long as nothing has been shown in the chat yet.
- Long tool names work. MCP tool names over 64 characters are shortened on the wire and restored on the way back, so OpenAI-format providers accept them.
- On Copilot Business/Enterprise, admins can disable third-party model providers via the "Bring Your Own Language Model Key" policy.
- Requires VS Code 1.104+ (older versions than 1.122 also need a signed-in Copilot plan — that's a VS Code rule, not ours).
Free & open source
OmniRoute is MIT-licensed and free forever — ⭐ star it on GitHub and join the project. This extension is MIT too; issues and PRs welcome at cheaper-inference/OmniCopilot. Maintainers publishing a new version: see docs/PUBLISHING.md (Marketplace + Open VSX).
© 2026 OmniRoute · A Cheaper Inference product · MIT License · Created by Diego Souza
OmniRoute is an independent open-source project, not affiliated with GitHub or Microsoft. GitHub Copilot is a trademark of GitHub, Inc.