AI Providers

Every AI backend you can use with MR Multi Model AI Coder, how to set each one up, and what it costs. Switch provider or model mid-conversation — the chat keeps running.

Provider	Auth	Cost	Best For
Claude	API key	Paid	Hard problems, long contexts
MR Coder	OAuth connect	With subscription	API key managed for you
Kilo Code	CLI OAuth	Free	Zero-config free tier
Qwen	API key or CLI	Free tier -> paid	Alibaba ecosystem
OpenRouter	API key	Free tier	Broadest free selection
Ollama	None	Free forever	Offline / private code
OpenAI Codex	API key or CLI	Paid (free trial)	OpenAI ecosystem, reasoning
Google Gemini	API key or CLI	Free tier -> paid	1M context, multimodal
HuggingFace	HF token	Free tier -> Pro	Open-weight agentic coders
Custom MCP (OpenAI)	API key (Bearer)	Depends on endpoint	Any /chat/completions gateway
Custom MCP (Anthropic)	API key (x-api-key)	Depends on endpoint	Any /v1/messages gateway

1. Claude (Anthropic API)

Anthropic's Agent SDK under the hood — full agentic loop, MCP servers, native checkpoints, prompt caching, extended thinking.

Model ID	Display Name	Context	Notes
`claude-opus-4-8`	Claude Opus 4.8	200K	Latest & most capable
`claude-opus-4-7`	Claude Opus 4.7	200K	Previous Opus
`claude-sonnet-4-6`	Claude Sonnet 4.6	200K	Recommended default
`claude-haiku-5-5`	Claude Haiku 5.5	1M	Fastest / cheapest

Setup

Get an API key from console.anthropic.com.

Ctrl+Shift+P -> "MR Multi coder: Set Claude API Key" -> paste.

(Optional) "Validate Claude API Key" to confirm.

Pick Claude in the provider chip of any chat.

Alternative auth: Set CLAUDE_CODE_USE_BEDROCK=1 for Amazon Bedrock, or CLAUDE_CODE_USE_VERTEX=1 for Google Vertex AI.

2. MR Coder (Router)

Claude models via the managed mr-coder.io service. No raw API key needed — connect your mr-coder.io account and the extension handles authentication automatically via OAuth.

Model ID	Display Name	Context	Notes
`claude-opus-4-8`	Claude Opus 4.8	200K	Latest & most capable
`claude-opus-4-7`	Claude Opus 4.7	200K	Previous Opus
`claude-sonnet-4-6`	Claude Sonnet 4.6	200K	Recommended default
`claude-haiku-5-5`	Claude Haiku 5.5	1M	Fastest / cheapest

Setup

Ctrl+Shift+P -> "MR Multi coder: Connect Account" to link your mr-coder.io account.

Approve the connection in your browser when prompted.

Pick MR Coder (Router) in the provider chip of any chat.

Models and rate limits are managed by your mr-coder.io subscription plan.

Full setup guide: MR Coder OAuth Connect

3. Kilo Code (CLI)

A single CLI install gives you access to 9 free models routed through Kilo's infrastructure. No API key, no per-model signups, no rate-limit dashboard.

Model ID	Display Name	Context
`kilo/kilo-auto/free`	Kilo Auto (Free)	128K
`kilo/x-ai/grok-code-fast-1:optimized:free`	Grok Code Fast	256K
`kilo/nvidia/nemotron-3-super-120b-a12b:free`	Nemotron 3 Super 120B	128K
`kilo/stepfun/step-3.5-flash:free`	Step 3.5 Flash	262K
`kilo/bytedance-seed/dola-seed-2.0-pro:free`	Dola Seed 2.0 Pro	128K

Setup

Bash

npm install -g @kilocode/cli

Install the CLI with npm install -g @kilocode/cli.

Ctrl+Shift+P -> "Sign In to CLI Provider" -> pick Kilo Code.

Follow the kilo login prompt — Ctrl-click the URL to complete auth.

Pick Kilo Code in the provider chip.

4. Qwen (DashScope API or CLI)

Alibaba's hosted API (DashScope) and ModelScope OAuth CLI.

Model ID	Display Name	Context
`qwen3-coder-plus`	Qwen3 Coder Plus	131K
`qwen3-coder-flash`	Qwen3 Coder Flash	131K
`qwen-max`	Qwen Max	32K
`qwen-plus`	Qwen Plus	131K
`qwen-turbo`	Qwen Turbo	128K

Setup — API mode

Get a DashScope key from dashscope.console.aliyun.com.

Ctrl+Shift+P -> "Set Qwen API Key" -> paste.

Pick Qwen in the provider chip.

Setup — CLI mode

Bash

npm install -g @qwen-code/qwen-code

Install the CLI with npm install -g @qwen-code/qwen-code.

Ctrl+Shift+P -> "Sign In to CLI Provider" -> pick Qwen.

Set mrMultiCoder.qwen.authMode to "cli" in settings.

5. OpenRouter

A single key gets you 16 curated free models plus access to every paid model OpenRouter hosts. No credit card required for the free tier.

Featured Free Models (fully agentic)

Model ID	Display Name	Context
`nvidia/nemotron-3-super-120b-a12b:free`	Nemotron 3 Super 120B	262K
`openai/gpt-oss-120b:free`	GPT-OSS 120B	131K
`openai/gpt-oss-20b:free`	GPT-OSS 20B	131K
`minimax/minimax-m2.5:free`	MiniMax M2.5	196K
`arcee-ai/trinity-large-preview:free`	Trinity Large Preview	131K

Other Free Agentic Models

Model ID	Display Name	Context
`qwen/qwen3-coder:free`	Qwen3 Coder 480B	262K
`google/gemma-4-31b-it:free`	Gemma 4 31B	262K
`meta-llama/llama-3.3-70b-instruct:free`	Llama 3.3 70B	65K

Free Tier Limits

- 20 requests/minute per model per key
- 50 requests/day without credits loaded
- 1000 requests/day after a one-time $10 load (never expires)

Setup

Generate a key at openrouter.ai/keys.

Ctrl+Shift+P -> "Set OpenRouter API Key" -> paste the sk-or-v1-... key.

Pick OpenRouter in the provider chip -> pick a model.

6. Ollama (Local)

Run models entirely on your own machine. No account, no API key, no rate limits, no internet required after download. Perfect for private codebases, offline work, and unlimited free iterations.

Model ID	Display Name	Disk	Context
`qwen3.6`	Qwen 3.6 35B	~24 GB	256K
`qwen3.6:27b`	Qwen 3.6 27B	~17 GB	256K
`qwen3-coder-next`	Qwen 3 Coder Next 80B	~52 GB	256K
`qwen2.5-coder:7b`	Qwen 2.5 Coder 7B (default)	~4.7 GB	32K
`qwen2.5-coder:3b`	Qwen 2.5 Coder 3B	~1.9 GB	32K
`llama3.3:70b`	Llama 3.3 70B	~40 GB	131K
`gemma3:27b`	Gemma 3 27B	~17 GB	131K
`deepseek-r1:32b`	DeepSeek R1 32B	~20 GB	131K
`phi4:14b`	Phi 4 14B	~9 GB	16K

RAM Guide

System RAM	Recommended	Runs Well
8 GB	`qwen2.5-coder:3b`	Quick edits, chat
16 GB	`qwen2.5-coder:7b`	Most day-to-day coding
32 GB	`qwen3.6:27b`	Longer tasks, 256K context
64 GB	`qwen3.6`	Flagship local models
128 GB+ or GPU	`qwen3-coder-next`	Heavy lifting

Install Ollama

Windows

Bash

# winget (built into Windows 10+)
winget install Ollama.Ollama

# Or download from https://ollama.com/download

macOS

Bash

brew install ollama

Linux

Bash

curl -fsSL https://ollama.com/install.sh | sh

Pull a Model & Use

Bash

# Pull the recommended model for 16 GB machines
ollama pull qwen2.5-coder:7b

# Or for 32 GB+
ollama pull qwen3.6:27b

Reload VS Code after installing the extension.

Open a chat -> click the provider chip -> pick Ollama (Local).

Pick a model you've pulled from the model chip.

Send a message.

Troubleshooting

Symptom	Fix
Ollama chip greyed out	ollama serve not running. Start it or check the tray icon.
"model not found"	Run ollama pull <tag> for the model you want.
First turn takes 30-60s	Model loading into RAM/VRAM. Subsequent turns are fast.
Out-of-memory crash	Pick a smaller tag (32b -> 14b -> 7b -> 3b).
Very slow on CPU	Normal for 7B+ models. Use :3b variants or add a GPU.

7. OpenAI Codex

OpenAI's models — GPT-4.1, o3, o4-mini, and codex-mini-latest — through either a direct API key or the official Codex CLI with OAuth.

Model ID	Display Name	Context
`o4-mini`	o4-mini (default)	200K
`o3`	o3	200K
`gpt-4.1`	GPT-4.1	1M
`gpt-4.1-mini`	GPT-4.1 Mini	1M
`gpt-4.1-nano`	GPT-4.1 Nano	1M
`codex-mini-latest`	Codex Mini (Latest)	200K

Setup — API mode

Get an API key from platform.openai.com/api-keys.

Ctrl+Shift+P -> "Set OpenAI Codex API Key" -> paste.

Pick OpenAI Codex in the provider chip.

Setup — CLI mode

Bash

npm install -g @openai/codex

Install the Codex CLI with npm install -g @openai/codex.

Ctrl+Shift+P -> "Log in with Codex CLI".

The extension auto-switches to CLI auth mode.

8. Google Gemini

Google's Gemini models with up to 1M token context windows. Generous free tier — no credit card required.

Model ID	Display Name	Context	Notes
`gemini-2.5-pro`	Gemini 2.5 Pro	1M	Most capable, deep reasoning
`gemini-2.5-flash`	Gemini 2.5 Flash	1M	Recommended default
`gemini-2.0-flash`	Gemini 2.0 Flash	1M	Fastest production model

Free Tier

- Gemini 2.5 Flash: 500 RPD, 15 RPM
- Gemini 2.5 Pro: 25 RPD, 5 RPM
- Gemini 2.0 Flash: 1500 RPD, 15 RPM

Setup — API mode

Get an API key from aistudio.google.com/apikeys (free, no credit card).

Ctrl+Shift+P -> "Set Gemini API Key" -> paste.

Pick Google Gemini in the provider chip.

Setup — CLI mode

Bash

npm install -g @google/gemini-cli

Install the CLI with npm install -g @google/gemini-cli.

Ctrl+Shift+P -> "Log in with Gemini CLI".

The extension auto-switches to CLI auth mode.

9. HuggingFace

22 curated open-weight agentic coding models via HuggingFace's OpenAI-compatible Inference API. Qwen, DeepSeek, Llama, Mistral, and more — all with a single HuggingFace token.

Model ID	Display Name	Context
`Qwen/Qwen3-235B-A22B`	Qwen3 235B A22B	131K
`Qwen/Qwen2.5-Coder-32B-Instruct`	Qwen2.5 Coder 32B (default)	131K
`deepseek-ai/DeepSeek-V3-0324`	DeepSeek V3	131K
`deepseek-ai/DeepSeek-R1-0528`	DeepSeek R1	131K
`meta-llama/Llama-4-Scout-17B-16E-Instruct`	Llama 4 Scout 17B	131K
`meta-llama/Llama-3.3-70B-Instruct`	Llama 3.3 70B	131K
`mistralai/Mistral-Small-24B-Instruct-2501`	Mistral Small 24B	32K
`microsoft/Phi-4-mini-instruct`	Phi-4 Mini	131K
`google/gemma-3-27b-it`	Gemma 3 27B	131K

Free Tier

- Anonymous: ~1 request/minute
- Free token: ~5 requests/minute
- Pro ($9/mo): Higher limits + priority routing
- Inference Endpoints: Dedicated GPU, no rate limits

Setup

Get a token at huggingface.co/settings/tokens (free). Enable "Make calls to the serverless Inference API".

Ctrl+Shift+P -> "Set HuggingFace API Token" -> paste.

Pick HuggingFace in the provider chip.

10. Custom MCP Chat Endpoints

Connect any endpoint that speaks one of two standard wire formats — it becomes a first-class provider in the chat, listed in the provider chip beside the built-ins.

Wire Format	Auth Header	Examples
OpenAI `/chat/completions`	`Authorization: Bearer`	Ollama, vLLM, LM Studio, LiteLLM, corporate gateways
Anthropic `/v1/messages`	`x-api-key`	Anthropic API, Bedrock/Vertex gateways, LiteLLM (Anthropic mode)

Setup via /mcp (recommended)

Open a chat and type /mcp -> Enter.

Pick OpenAI Compatible Endpoint or Anthropic Compatible Endpoint.

Enter the endpoint URL (just the base, e.g. https://api.anthropic.com).

Enter the API key — stored in SecretStorage, never in settings.json.

Enter the model id the endpoint serves (e.g. claude-sonnet-4-6, gpt-4o).

The provider appears in the chip immediately — no reload needed.

Setup via MCP modal

Click the MCP button in the chat toolbar.

Pick a preset: OpenAI Compatible or Anthropic Compatible.

Fill in URL, API key, and model id -> Save.

Troubleshooting

Symptom	Fix
401 / 403 on first message	API key is wrong or expired. Run /mcp to reconnect.
404 on first message	Wrong model id. Run /mcp and enter an id the endpoint serves.
Provider not in chip after save	Upgrade to v0.7.6+ (fixed: pushState now fires after save).

Mix & Match — Switch Mid-Chat

Every provider chip is live. Click a different one at any time — the session tears down, a fresh one spawns, and prior messages replay automatically.

Claude for design, Ollama for implementation — sketch the approach on Claude, switch to Ollama for free iterations.

OpenRouter when Claude is rate-limited — flip the chip when you hit Claude's daily limit.

Kilo for edits, Qwen for refactors — the Kilo auto-router handles small patches; swap to Qwen for thinking-heavy rewrites.

FAQ

Do I need a credit card?

No. Claude and OpenAI Codex require paid credit, but Kilo, OpenRouter, Ollama, Gemini, and HuggingFace all have free tiers with no card required.

Which is the "always works" option?

Ollama. It's on your own machine — no rate limits, no account flagging, no API outages.

Which free model stays responsive?

For quick work: Nemotron 3 Super 120B or GPT-OSS on OpenRouter. For heavy tasks: Ollama local.

Where are my keys stored?

In your OS secure keychain (Keychain on macOS, Credential Manager on Windows, libsecret on Linux). Never in settings.json.