GitHub Copilot Custom Endpoints

TL;DR — GitHub Copilot switched to usage-based billing on June 1, 2026. Every chat and agent session now burns AI credits — fast. This repo shows you how to plug cheaper non-GitHub models (DeepSeek, Kimi, Qwen, MiMo, MiniMax, GLM) into VS Code's Copilot chat — often 5–55× cheaper than the built-ins — while keeping agent mode, tools, streaming, and vision.

What is this?

VS Code lets you add your own language-model endpoint via a small JSON config file. Many providers advertise "OpenAI-compatible" APIs but reject the exact request shapes VS Code sends. This repo collects real, tested setups — one per provider — plus a tiny local proxy that smooths over the rough edges when needed.

If OpenRouter is blocked by your network or too generic for your model's quirks, this is the workaround.

How it works (5 steps)

1. Pick a model from the table below.
2. Add it to your VS Code config — copy the snippet from the model's doc.
3. Set the API key through VS Code's UI (it goes to your OS keychain, not the file).
4. Configure the Utility Small Model — VS Code now requires a fast fallback model for built-in utility flows. Open Settings → search "Chat: Utility Small Model" → pick your fastest custom-endpoint model (or "Default" if you have native Copilot models available). Without this, chat may not function correctly.
5. Open chat and pick the model from the model picker.

That's it. No code, no servers to manage (unless the model specifically needs the local proxy — the table tells you).

Pick a model

| Model | Provider | Needs proxy? | Vision | Setup guide | | ---------------------------- | ---------- | ----------------------- | ------ | ----------------------------------------------------------------------------------------------------------------------------------------- | | Kimi K3.1 Pro | Moonshot | Yes | ✅ | Manual setup | | Kimi K3 | Moonshot | Yes | ✅ | Manual setup | | MiMo V2.6 Pro | Xiaomi | Yes | ✅ | Extension / Manual setup | | MiMo V2.6 Flash | Xiaomi | Yes | ✅ | Extension / Manual setup | | MiMo V2.6 Pro UltraSpeed | Xiaomi | Yes | ✅ | Extension / Manual setup | | Qwen 3.8 Max (0803) | OpenRouter | Optional (experimental) | ✅ | Snapshot/proxy setup | | Qwen 3.8 Max (0902) | DashScope | Optional (recommended) | ✅ | Manual setup | | MiniMax M3.1 Flash | MiniMax | No | ✅ | Manual setup | | MiniMax M3 | MiniMax | No | ✅ | Extension ★ / Manual setup | | GLM 5.3 Flash | Z.ai | No | ✅ | Manual setup | | GLM 5.3 | Z.ai | No | ❌ | Manual setup | | DeepSeek V4.1 Flash 1017 | DeepSeek | No (uses an extension) | ✅ | Extension / setup notes |

Setup

1. Find (or create) your config file

| OS | Path | | ------- | ----------------------------------------------------------------- | | Windows | %APPDATA%\Code\User\chatLanguageModels.json | | macOS | ~/Library/Application Support/Code/User/chatLanguageModels.json | | Linux | ~/.config/Code/User/chatLanguageModels.json |

If the file doesn't exist yet, create it with [] inside.

2. Add a model entry

Open the setup guide for the model you picked (links in the table above) and copy its JSON snippet into the file. Each snippet is a single provider object inside the array.

⚠️ Leave apiKey as "" — never paste the key into the JSON file.

3. Set the API key

1. Open the Command Palette (Ctrl+Shift+P / Cmd+Shift+P).
2. Run Chat: Manage Language Models.
3. Find your provider in the list, right-click the group name → Update API Key.
4. Paste your key. It's stored in your OS keychain.

4. Configure the Utility Small Model

⚠️ Required for BYOK/custom-endpoint users. VS Code's latest update mandates that you explicitly set which model handles built-in small/fast utility flows. Without this, custom-endpoint models may not function correctly in chat or agent mode.

1. Open VS Code Settings (Ctrl+, / Cmd+,).
2. Search for "Chat: Utility Small Model" (setting ID: chat.lm.utilitySmallModel).
3. Pick the fastest, cheapest model available to you from the dropdown. Good choices:
   • GLM 5.3 Flash — if you have GLM configured (cheapest custom-endpoint option, ~$0.13/session).
   • GPT-6 Luna — if it is available through your native Copilot models (fast, ~$0.10/session).
   • MiMo V2.6 Flash — if you already have Xiaomi MiMo configured (low-cost multimodal option, ~$0.10/session).
   • Default — if you still have native Copilot model access, this lets VS Code use its built-in fast model.
4. The setting takes effect immediately — no restart needed.

Why this matters: VS Code uses a small utility model for quick background tasks (token counting, prompt truncation, lightweight completions). When you switch to custom-endpoint models, the framework still needs a fast model for these utility flows. If left unset, some features may silently degrade or fail.

5. Chat

• Open Copilot chat (Ctrl+Alt+I / Cmd+Ctrl+I).
• Click the model picker (top-right).
• Pick your model and ask something.

If a model needs a proxy, the setup guide will tell you to run a command first. Keep that terminal open while you chat.

Common commands

Run from the repo root:

npm run proxy        # Start all proxies (Kimi + Qwen + MiMo)
npm run proxy:kimi   # Start only the Kimi proxy
npm run proxy:qwen   # Start only the Qwen proxy
npm run proxy:mimo   # Start only the MiMo proxy
npm run clean:logs   # Remove debug_log/
npm test             # Run the test suite

Or globally via npx (no clone needed):

npx copilot-custom-endpoint          # Start all proxies
npx copilot-custom-endpoint kimi     # Kimi only
npx copilot-custom-endpoint qwen     # Qwen only
npx copilot-custom-endpoint mimo     # MiMo only
npx copilot-custom-endpoint clean    # Remove debug_log/

Pricing snapshot

All prices are USD per 1M tokens (non-cached). 1 AI credit = $0.01; multiply a dollar amount by 100 to convert it to credits (for example, $5.00/1M = 500 credits/1M).

The table is sorted by Cost per intelligence, lowest first. This is a value metric calculated as:

estimated session cost / Intelligence Score

It approximates the cost of one point on the AA Intelligence Index under the common workload below. A lower number means more benchmark intelligence per estimated dollar; it is not an overall quality ranking or a provider billing quote.

Est. session uses ~10K input + ~2K output tokens per turn over 50 turns: 500K input tokens plus 100K output tokens total. For each model, the estimate is:

(0.5 × input price) + (0.1 × output price)

For example, GLM 5.3 Flash is (0.5 × $0.15) + (0.1 × $0.50) = $0.125, shown as ~$0.13. Cached-input discounts are not included. Actual costs vary with context length, output length, caching, pricing tiers, promotions, and provider billing. DeepSeek estimates use peak rates; official off-peak rates are half.

The current AA values come from OpenRouter's model metadata and ranking card; every pricing cell comes from first-party provider pricing or GitHub Copilot's official billing table, never OpenRouter. The OpenRouter ranking card places unversioned Qwen3.8 Max at 53.4 (#3), while the exact 0902 API metadata row is 45.4. Qwen Cloud PAYG pricing remains $2 / $0.25 implicit cache / $6 for Qwen3.8 Max; its separate Token Plan is not the PAYG rate table. This snapshot records Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.7, GPT-6 Sol, GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5 as Copilot-native models, using current AA v4.3.2 composites. Gemini 4 Argon (released October 13, 2026) enters at 53 on Google's introductory pricing of $2 / $0.10 implicit cache / $10 per 1M tokens, which doubles to $4 / $0.20 / $20 on December 1, 2026. Claude Haiku 5.5 (released October 7, 2026) enters at 43 and, at ~$0.0023 per intelligence point, becomes the lowest cost per intelligence in the table; MiMo V2.6 Flash (37.9) remains the cheapest custom-endpoint option. Muse Spark 1.3 (Meta, released September 2, 2026) enters at 48 on Meta Model API pricing of $1.25 / $0.15 cached input / $4.25 per 1M tokens; its day-one muse-spark-1.3-contributor SKU is excluded from this snapshot because Meta may train future models on those prompts and completions. DeepSeek's V4.1-Flash-1017 build (released October 17, 2026) replaces the September 10 build under the same deepseek-flash API ID: its composite rises from 39.5 to 45 at unchanged peak rates ($0.30 / $0.006 / $1.20), moving it to ~$0.0060 per intelligence point between GPT-5.6 Luna and MiMo V2.6 Pro. MiniMax's MiniMax-M3.1-Flash (announced October 19, 2026) enters at 43 on flat pricing of $0.15 / $0.015 cache read / $0.60 per 1M tokens — no prompt-length tiers and no service_tier options — which puts it at ~$0.0031 per intelligence point between GLM 5.3 Flash and GPT-5.6 Luna. Artificial Analysis measures it at 96 tokens/s and very verbose (180M output tokens to run the index vs a 100M median), so real sessions can cost more than the modeled ~$0.14.

| Model | Provider | Cost per intelligence | Intelligence Score | Est. session | Vision | Context window | | ---------------------------- | --------- | --------------------- | ------------------ | ------------ | ------ | -------------- | | Claude Haiku 5.5 | Anthropic | ~$0.0023 | 43 | ~$0.10 | ✅ | 1M | | MiMo V2.6 Flash | Xiaomi | ~$0.0026 | 37.9 | ~$0.10 | ✅ | 1M | | GPT-6 Luna | OpenAI | ~$0.0027 | 37.3 | ~$0.10 | ✅ | 1M | | GLM 5.3 Flash | Z.ai | ~$0.0030 | 41.8 | ~$0.13 | ✅ | 1M | | MiniMax M3.1 Flash | MiniMax | ~$0.0031 | 43 | ~$0.14 | ✅ | 1M | | GPT-5.6 Luna | OpenAI | ~$0.0059 | 37.3 | ~$0.22 | ✅ | 1M | | DeepSeek V4.1 Flash 1017 | DeepSeek | ~$0.0060 | 45 | ~$0.27 | ✅ | 1M | | MiMo V2.6 Pro | Xiaomi | ~$0.0066 | 46.3 | ~$0.30 | ✅ | 1M | | MiniMax M3 | MiniMax | ~$0.0093 | 29.2 | ~$0.27 | ✅ | 1M | | Gemini 3.8 Flash | Google | ~$0.0183 | 40.9 | ~$0.75 | ✅ | 1M | | Muse Spark 1.3 | Meta | ~$0.0219 | 48 | ~$1.05 | ✅ | 1M | | GLM 5.3 | Z.ai | ~$0.0254 | 44.8 | ~$1.14 | ❌ | 1M | | Qwen 3.8 Max (0803) | DashScope | ~$0.0300 | 53.4 | ~$1.60 | ✅ | 1M | | Grok 4.7 | xAI | ~$0.0345 | 46.4 | ~$1.60 | ✅ | 500K | | Qwen 3.8 Max (0902) | DashScope | ~$0.0352 | 45.4 | ~$1.60 | ✅ | 1M | | Claude Sonnet 5.5 | Anthropic | ~$0.0357 | 56.0 | ~$2.00 | ✅ | 1M | | Grok 4.6 | xAI | ~$0.0361 | 44.3 | ~$1.60 | ✅ | 500K | | Kimi K3.1 Pro | Moonshot | ~$0.0363 | 55.1 | ~$2.00 | ✅ | 1M | | Gemini 4 Argon | Google | ~$0.0377 | 53 | ~$2.00 | ✅ | 1M | | GPT-6.1 Sol | OpenAI | ~$0.0386 | 51.8 | ~$2.00 | ✅ | 1M | | GPT-6 Sol | OpenAI | ~$0.0421 | 47.5 | ~$2.00 | ✅ | 1M | | GPT-5.6 Terra | OpenAI | ~$0.0523 | 42.1 | ~$2.20 | ✅ | 1M | | Claude Sonnet 5 | Anthropic | ~$0.0524 | 38.2 | ~$2.00 | ✅ | 1M | | MiMo V2.6 Pro UltraSpeed | Xiaomi | ~$0.0658 | 46.3 | ~$3.05 | ✅ | 1M | | Kimi K3 | Moonshot | ~$0.0688 | 43.6 | ~$3.00 | ✅ | 1M | | Claude Opus 5.5 | Anthropic | ~$0.0694 | 57.6 | ~$4.00 | ✅ | 1M | | GPT-5.6 Sol | OpenAI | ~$0.0851 | 47.0 | ~$4.00 | ✅ | 1M | | Claude Opus 5 | Anthropic | ~$0.0984 | 50.8 | ~$5.00 | ✅ | 1M | | Claude Fable 5.1 | Anthropic | ~$0.1873 | 53.4 | ~$10.00 | ✅ | 1M | | GPT-6 Astra | OpenAI | ~$0.1898 | 52.7 | ~$10.00 | ✅ | 1M | | Claude Fable 5 | Anthropic | ~$0.2016 | 49.6 | ~$10.00 | ✅ | 1M |

MiMo V2.6 Pro UltraSpeed uses the MiMo V2.6 Pro benchmark score because it is the same-quality latency tier; no independent composite is published. MiniMax M3.1 Flash is a separate model with its own flat rates; MiniMax's API rejects service_tier for it. Kimi K3.1 Pro supersedes K3 at a lower price ($2 / $0.20 / $10 vs $3 / $0.30 / $15 per 1M); K3 remains served. MiMo V2.6 Flash's composite (37.9) is OpenRouter's Artificial Analysis v4.3.2 value; Xiaomi's published launch benchmarks for it are recorded in docs/benchmarks.md. GPT-6 Sol, GPT-6.1 Sol, and Luna use standard rates up to 272K input tokens; requests above that threshold have higher input/cache and output rates. GitHub Copilot also bills cache writes separately. Plan access differs: GPT-6 Sol and GPT-6.1 Sol require Pro+ or higher, Luna and Claude Sonnet 5.5 are available from Pro, Claude Opus 5.5 requires Pro+ or higher, and Claude Haiku 5.5 is available on every paid plan (Pro through Enterprise). Gemini 4 Argon is Copilot-native for Pro+, Max, Business, and Enterprise (not Pro); Google's $2 / $0.10 implicit cache / $10 per 1M introductory pricing (output includes thinking tokens, implicit caching only) is guaranteed only through November 30, 2026 and doubles to $4 / $0.20 / $20 on December 1, 2026. Claude Haiku 5.5 is the only current Claude model priced by prompt length: $0.10 / $0.01 cache read / $0.50 per 1M for prompts up to 100,000 tokens, and $0.50 / $0.05 / $2.50 above that. The table shows the sub-100K rates, which Anthropic says cover ~90% of Haiku-class requests. Muse Spark 1.3 is closed and API-only on the Meta Model API (muse-spark-1.3); the 12–21× cheaper muse-spark-1.3-contributor SKU is excluded from this snapshot because Meta may use prompts and completions to train future models. Meta reports audio understanding is degraded versus 1.2.

Detailed pricing, benchmark provenance, and model-specific notes live in docs/pricing.md and docs/benchmarks.md. This page stays focused on setup and model selection. For a copy-paste config containing all providers at once, see docs/example-config.md.

👤 Personal picks —

• Claude Haiku 5.5 — the lowest cost per intelligence in this table (~$0.0023) and the best value among Copilot-native models, at about $0.10/session with vision.
• MiMo V2.6 Flash — the lowest cost per intelligence among custom-endpoint models (~$0.0026), at about $0.10/session with vision.
• GLM 5.3 Flash — the runner-up among custom-endpoint models (~$0.0030), at about $0.13/session with vision.

Companion tools

ℹ️ These are third-party tools — not built into the custom endpoints or proxies in this repo. Each one must be installed, configured, and (where applicable) billed for separately, directly with its own provider. Nothing here is bundled, proxied, or auto-configured by copilot-custom-endpoint or the per-model setups above. The entries below are just pointers to tools the author has found useful alongside the model configs.

These work alongside the providers above and fill gaps that VS Code's built-in tool surface doesn't cover natively.

🎬 Video Context MCP — video understanding for AI coding assistants

VS Code's built-in view_image tool only accepts static images (PNG, JPG, GIF, WebP). That's a hard wall if you want to ask an AI assistant about a screen recording, a screencast, a product demo, or any other video. Several vision-capable models in this repo actually accept video natively — but VS Code's tool pipeline never gets the chance to forward it.

Video Context MCP is a small MCP server that bridges that gap. It works with GitHub Copilot, Cursor, and Claude Code out of the box, and:

• Extracts frames from local files or remote URLs (no ffmpeg gymnastics required).
• Routes them through a multi-provider fallback chain — Gemini → GLM 4.6V Flash → Qwen3.8-max.
• Answers natural-language questions about the video grounded in actual frames: "what does the speaker click in the last 30 seconds?", "summarize the demo", "find the frame where the error appears".
• Extras: timestamp search, audio transcription with speaker diarization, and video metadata (resolution, duration, codec).

🪣 Bitbucket MCP — secure, read-only Bitbucket access for VS Code Copilot, Cursor, and Claude Code

GitHub ships a first-party MCP server (and it's even bundled into Copilot), so asking "what's open in my org's repos / show me PR #123" works seamlessly on github.com. Bitbucket has no equivalent — Atlassian hasn't shipped one — which leaves Bitbucket Cloud users copy-pasting PR URLs, diffs, and file contents into chat by hand.

Bitbucket MCP is a small MCP server that closes that gap. It works with VS Code GitHub Copilot, Cursor, and Claude Code out of the box, and:

• 38 tools across 8 categories — repositories, pull requests, branches & commits, diffs & comparisons, CI/CD pipelines, issues, code search, and users.
• Read-only by design — makeRequest() blocks all non-GET requests at runtime, so no write, delete, or modify operation is possible.
• Token-friendly output — every tool supports text, json, and toon (Token-Oriented Object Notation) formats; toon cuts LLM token consumption by 30–60% on large PR/commit lists.
• JMESPath filtering on all 38 tools, so you can trim responses (e.g. only open PRs, or just title + author) before they hit the model.
• One-call PR context — bb_get_context bundles PR metadata, diffstat, CI statuses, and comments in a single request.
• Drop-in install — npm install -g @tugudush/bitbucket-mcp plus a short .vscode/mcp.json entry, authenticated with a Bitbucket API token + your Atlassian email.

Need help?

• Per-model issues: check the troubleshooting section at the bottom of each model's doc.
• Repo questions / bugs: open an issue on GitHub.

Repo layout

.
├── docs/models/<provider>-<model>.md   # Per-model setup guides (the real docs)
├── proxy/                              # Local compatibility shims
├── tests/                              # Test assets
└── debug_log/                          # Runtime logs (git-ignored)

Want to add a new model?

1. Create docs/models/<provider>-<model>.md with a clear walkthrough.
2. If the provider needs request rewriting, add a proxy under proxy/.
3. Submit a PR.

The current AA values come from OpenRouter's model metadata and ranking card; every pricing cell comes from first-party provider pricing or GitHub Copilot's official billing table, never OpenRouter. The OpenRouter ranking card places unversioned Qwen3.8 Max at 53.4 (#3), while the exact 0902 API metadata row is 45.4. Qwen Cloud PAYG pricing remains $2 / $0.25 implicit cache / $6 for Qwen3.8 Max; its separate Token Plan is not the PAYG rate table. This snapshot records Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.7, GPT-6 Sol, GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5 as Copilot-native models, using current AA v4.3.2 composites.

• Chat only. Inline completions, semantic search, and next-edit suggestions still need a GitHub-hosted model.
• Each proxy is tuned for a specific provider family. Don't point the Kimi proxy at an arbitrary OpenAI-compatible endpoint.

Support

If this helped, consider sponsoring or donating:

Solana (SOL)

CWZccD3Ny3XotFZtnkcyzP3hapmu3ExknN1PF4rEvP3u