MR Multi Model Ai coderMR Coder — 8 AI providers in one extension. Claude, Kilo Code (free), OpenRouter (14 free models), Ollama (offline, Qwen 3.6 / Llama / Gemma), Qwen, OpenAI Codex, Google Gemini, HuggingFace (Qwen / DeepSeek / Llama / Mistral). Switch mid-chat, bring your own keys or use completely free models. Installation Launch VS Code Quick Open ( Ctrl+P), paste the following command, and press enter.Copied to clipboard |
MR Multi Model Ai CoderThe Claude Code experience — on your own keys, your own CLIs, a fully-managed SaaS option, a tunneled remote web instance, an optional Product Owner agent, and a pile of completely free models. 👉 Install from the VS Code Marketplace 📚 Providers · Features · Commands · User documentation · Architecture · Web Remote Access · Analytics
Ever been locked out of Claude in the middle of your work?You're deep in a problem. Claude is helping you ship. Then — nothing. You're locked out.
Whatever the reason, your whole day stops while you figure out what to do next. You lose your train of thought. You lose the conversation you were halfway through. And the problem you were about to solve is still sitting there. This extension makes that problem go away. Use your own Claude account when it's working. The moment it isn't — click once and switch to a completely free model. Your chat keeps going. The files you had open, the context, the history, everything comes along. You never lose your flow. You get ten provider chips in one extension:
Plus two more paths that ship inside the same chips:
Mix and match however you want. Pay nothing you don't need to. Full setup, model tables, and config walkthroughs for every provider in PROVIDERS.md. Free models, out of the boxThree separate free paths. No credit card on any of them. Kilo Code (
OpenRouter (sign up free at openrouter.ai/keys — no credit card)
Ollama Local (install,
Qwen — DashScope free trial (API) or the Qwen CLI (ModelScope OAuth) Mix and match. Use Claude for the hard reasoning, flip to a free model for routine edits, pay for nothing you don't need. iFlow support is currently paused while the iFlow CLI's sign-in flow stabilises — it'll come back once their auth migration is complete. What you getEverything Claude Code users expect, plus a few things they don't:
Haptic Feedback & Progress MusicEvery chat event has a sound cue — submit click, response ding, error thud, approval tone, task-complete chime. All work out of the box with built-in synthesized tones. You can replace any of them with your own audio files ( Progress MusicPlay background music while the AI is generating:
Files play sequentially and loop until the response finishes. A single file loops continuously. Volume is independent from event sounds ( Multiple tabs? Only one tab plays at a time. When that tab finishes, music automatically resumes in the next still-generating tab. Reset everything: Full settings reference and examples in COMMANDS.md § Haptic Feedback. InstallEasiest — VS Code:
Command line: Marketplace page: https://marketplace.visualstudio.com/items?itemName=mr-innovations.mr-multi-coder Quick start
Every conversation is a standalone editor tab. Open five in parallel if you want. ProvidersTen live provider chips, one hidden, plus two CLI auth paths (Qoder and
Key commandsOpen the Command Palette ( Chat & sessions
Providers & auth
Web & tunnel
Code context
MCP
Product Owner (when enabled — see FEATURES.md)
Analytics & diagnostics
Haptic feedback
Slash commands inside the chat: Full command reference with keybindings, haptic feedback settings, and slash commands: COMMANDS.md. Scheduled messagesType
If VS Code is closed when the scheduled time passes, the message is not auto-sent. Instead, a notification shows you what was missed and offers three actions: Send Now, Reschedule, or Delete. Key settings (under
|
| Setting | Default | What it does |
|---|---|---|
provider |
claude |
Active backend (claude, mrcoder, kilo, qwen, openrouter, ollama, codex, gemini, huggingface) |
permissionMode |
default |
default / plan / acceptEdits / bypassPermissions |
claude.model |
claude-sonnet-4-6 |
Model for the Claude provider |
claude.backend |
anthropic |
Anthropic / Bedrock / Vertex / Foundry |
claude.credentials[] |
[] |
Multi-account credential profiles (custom endpoints, custom headers) |
claude.activeCredentialId |
"" |
Which profile is active for new sessions |
claude.extendedThinking |
false |
Enable Claude's server-side reasoning |
claude.thinkingBudgetTokens |
32000 |
Token budget for extended thinking |
mrcoder.model |
claude-sonnet-4-6 |
Default model when on the MR Coder chip |
qoder.cliPath |
qodercli |
Path to Qoder CLI binary |
qwen.authMode |
api |
api (DashScope) or cli |
openrouter.model |
nvidia/nemotron-3-super-120b-a12b:free |
Default OpenRouter model |
ollama.endpoint |
http://localhost:11434/v1 |
Your Ollama server URL |
ollama.model |
qwen2.5-coder:7b |
Default Ollama tag |
codex.authMode |
api |
api (OpenAI key) or cli (Codex CLI OAuth) |
codex.model |
o4-mini |
Default OpenAI model |
gemini.authMode |
api |
api (AI Studio key) or cli (Gemini CLI OAuth) |
gemini.model |
gemini-2.5-flash |
Default Gemini model |
huggingface.endpoint |
https://router.huggingface.co/hf-inference/v1 |
HF Inference API URL (override for dedicated endpoints) |
huggingface.model |
Qwen/Qwen2.5-Coder-32B-Instruct |
Default HF model |
maxTurns |
1000 |
Hard cap on agent turns per session |
maxAutoContinue |
5 |
Auto-resend "Continue" this many times when truncated |
maxToolRounds |
0 |
Max tool-call rounds per turn; 0 = provider default |
customSystemPrompt |
"" |
Appended to every system prompt |
autoCompactThreshold |
0.7 |
Trigger auto-compact at this fraction of context |
sizeCompactThreshold |
2000 |
Trigger auto-compact when stored messages exceed this count |
contextStrategy |
rolling-summary |
truncate (drop oldest) or rolling-summary (LLM-summary + last N raw) |
rollingSummary.summarizeEveryNMessages |
25 |
New user messages before summariser fires |
rollingSummary.keepRawAfterSummary |
10 |
Recent user turns kept verbatim post-summary |
rollingSummary.summariserProvider |
auto |
auto (cheapest available) / claude / active |
swapSummarise |
auto |
LLM-summarise when swapping provider? auto / always / never |
contextReplay.{maxChars,maxMessages,perMessageCap} |
80000 / 50 / 4000 |
Replay-prefix budget on resume / provider swap |
toolResultTruncateBytes |
50000 |
Cap large tool results |
loadProjectMemory |
true |
Auto-load CODER.md / CLAUDE.md / AGENTS.md / .mrcoder/MEMORY.md |
loadGlobalMemory |
true |
Also auto-load ~/.mrcoder/MEMORY.md |
showMemoryInStatusBar |
true |
Status-bar chip with loaded project-memory size (click to inspect) |
sessionMemory.maxSummaries |
3 |
Compact summaries kept in per-session MEMORY.md |
costCapUsdPerSession |
0 |
Block sends past this cost; 0 disables |
showCostInStatusBar |
true |
Show running token usage / cost in status bar (Claude / MR Coder) |
telemetry.enabled |
false |
Opt-in only, no content / paths / keys |
autoAttachSelection |
false |
Auto-attach active editor selection on send |
enableNewConversationShortcut |
false |
Ctrl+N in sidebar = new chat |
mcpServers |
{} |
Same JSON shape as Claude Code's mcpServers |
analytics.enabled |
true |
Toggle analytics events |
analytics.refreshIntervalSeconds |
30 |
Dashboard auto-refresh interval |
analytics.minutesPerTask |
5 |
"Time Saved" KPI multiplier |
analytics.goals.* |
0 |
Acceptance-rate / daily-prompt / cost / response-time goal lines |
po.enabled |
false |
Turn on the Product Owner agent |
po.mode |
passive |
passive posts notes; active auto-injects the next task |
po.autoCommit |
true |
Auto-commit the AI's work as MR CODER PO (when PO is enabled) |
po.snapshotViewEnabled |
true |
Show the PO Snapshots view |
po.endpoint / po.endpointToken |
"" |
Optional upload URL + bearer for sharing snapshots |
webAccess.* |
varies | Web remote access (port, projectsFolder, autoStart, allowedExtensions, sessionTimeout, maxSessions) |
tunnel.apiBaseUrl |
https://mr-coder.io |
Tunnel API base URL (override for self-hosted) |
tunnel.autoConnect |
true |
Auto-start cloudflared when VS Code opens |
hapticFeedback.* |
varies | Sound events, keyboard sounds, progress music |
rollout.killSwitch |
false |
Operator kill-switch (off by default — flip in a hotfix bundle) |
Team conventions (/init)
Run /init in any chat. It asks you a handful of questions (max file size, testing policy, commit style, off-limits paths, etc.), writes a single CODER.md at the repo root, and wires a reference to it into CLAUDE.md, KILO.md, IFLOW.md, QWEN.md, CODEX.md, GEMINI.md, and AGENTS.md so every provider reads the same conventions.
Change the rules later? Edit CODER.md — everything else picks it up automatically.
Web Remote Access
Access your workspace from any browser — phone, tablet, or another computer — with full authentication and per-user access control.
- Open MR Multi coder Settings → Web Remote Access section
- Add users (email + password), assign project access (all or specific folders)
- Click Start Server — a login-protected VS Code Web instance launches
What you get:
- Two-port architecture — auth proxy on public port (18462), VS Code Server on internal-only port (18463). No unauthenticated traffic reaches the server.
- JWT session auth with HttpOnly cookies, bcrypt password hashing, brute-force rate limiting
- Per-user project access — restrict users to specific workspace folders
- Extension allowlist — only MR Multi coder is installable in the web instance (uses VS Code 1.96+
extensions.allowed) - Auto-installs MR Multi coder into the web VS Code instance so remote users get the full AI coding experience
- Works on LAN, VPN, or with a reverse proxy for public access
Full setup, security model, and troubleshooting in REMOTE-ACCESS.md.
Security
- API keys live in VS Code's
SecretStorage— never insettings.json, never synced, never logged. - CLI providers are auto-detected on your PATH. Nothing is bundled.
- Untrusted / virtual workspaces correctly gate file-edit and shell tools.
- Workspace trust respected throughout — Bash and edit tools refuse to run in untrusted folders.
- Web Remote Access uses JWT + bcrypt + rate limiting — see REMOTE-ACCESS.md for the full security model.
Org policy (optional)
Commit a .mrcoder/policy.json to your repo to lock behaviour for contributors:
{
"forcedPermissionMode": "plan",
"forcedDisallowedTools": ["Bash"],
"costCapUsdPerSession": 5,
"disableCliProviders": false,
"lockProvider": "claude"
}
Documentation
- PROVIDERS.md — every provider in one place: model tables, free-tier caveats, Ollama install + configure walkthrough, switching mid-chat, FAQ.
- FEATURES.md — complete catalog of every feature, every slash command, every setting.
- COMMANDS.md — every command, keybinding, slash command, and haptic feedback setting with examples.
- DOCUMENTATION.md — day-to-day user guide: install, configure providers, attachments, rewind, troubleshooting.
- ARCHITECTURE.md — how the extension is built: bundle layout, protocol, per-tab lifecycle, security posture.
- ANALYTICS_FEATURES.md — developer analytics dashboard: productivity score, streak tracking, cost intelligence, model ranking, export.
- REMOTE-ACCESS.md — web remote access: setup, security model, per-user project access, troubleshooting.
- CHANGELOG.md — what changed in each release.
Feedback
Bug, missing feature, weird edge case? File it on the Marketplace page, join the Telegram community, or reach out to the publisher.
- Telegram community: https://t.me/+Bj5YpMhXYHwzNzc1 — fastest route for questions, issue reports, and suggestions.
- Marketplace: https://marketplace.visualstudio.com/items?itemName=mr-innovations.mr-multi-coder
- Publisher: MR Innovations
Legal
MIT licensed — see LICENSE.
Not affiliated with Anthropic. "Claude Code" is Anthropic's trademark. This extension is independent and uses Anthropic's official Agent SDK for the Claude provider. "Kilo Code", "iFlow", and "Qwen" are independent CLIs from their respective authors.