OAIProxy

A self-maintained VS Code extension to use Azure Foundry/OpenAI/Ollama/Anthropic/Gemini API providers in GitHub Copilot Chat, with presets for Azure Foundry, OpenAI, TokenRouter, Anthropic, Fireworks, Kimi, DeepSeek, Z.AI GLM, Xiaomi MiMo, and MiniMax 🔥

English | 简体中文

Features

• Multi-API support: Azure Foundry/OpenAI/Ollama/Anthropic/Gemini APIs, with direct Azure Foundry presets plus OpenAI-compatible presets for Fireworks, Kimi, DeepSeek, Z.AI GLM, Xiaomi MiMo, MiniMax, ModelScope, SiliconFlow, and more
• Vision models: Full support for image understanding capabilities
• Vision Bridge: Use images in chat with text-only models — OAIProxy automatically describes images via a configured vision model with LRU caching
• Think tag support: Seamless display of model thinking/reasoning blocks across all providers (OpenAI, Ollama, Gemini, Anthropic)
• Thinking Effort control: VS Code's built-in per-model Thinking Effort dropdown in the model picker — customize reasoning effort on the fly
• Advanced configuration: Flexible chat request options with thinking/reasoning control
• Multi-provider management: Configure models from multiple providers simultaneously with automatic API key management
• Model connection tests: Test one configured model or run up to four tests in parallel from Model Management using the model's real saved request configuration
• Provider usage checks: Check xAI/Grok weekly subscription credit remaining, OpenAI Codex OAuth quota windows, Fireworks month-to-date serverless tokens, DeepSeek/Kimi credit balance, TokenRouter Management Key credit balance, MiniMax token-plan remaining quota, and OpenAI/Anthropic month-to-date cost usage from a standalone Provider Usage Check table; Azure Foundry, Z.AI, and MiMo explain where no public inference-key usage endpoint is available
• Multi-config per model: Define different settings for the same model (e.g., GLM-4.6 with/without thinking)
• Visual configuration UI: Intuitive interface for managing providers and models
• Auto-retry: Handles API errors (429, 500, 502, 503, 504) with exponential backoff
• Request cancellation: Stop in-progress chat requests instantly — cancellation is wired to HTTP AbortController across all API modes
• Token usage: Shared text/image/tool accounting, a prepared-request context guard, and optional usage calibration; see the Token Estimation Guide
• Live output speed: Estimated tokens per second beside the status-bar context percentage, updated every 500 ms; hover for output counts, duration, and the final response average.
• Git integration: Generate commit messages directly from source control with OpenAI/OpenAI Responses/Ollama/Anthropic models
• Import/export: Easily share and backup configurations
• Tools optimization: Optimize agent read_file tool handling for supported streamed tool calls, avoiding small chunks for large files.
• Structured logging: File-based request/debug logs with rotation, configurable levels (off/debug/info/warn/error)

Requirements

• VS Code 1.120.0 or higher.
• OpenAI-compatible provider API key, or an eligible ChatGPT/Codex account for the OpenAI OAuth cards.

Quick Start

1. Install the OAIProxy VSIX package (lqdflying.oaiproxy).
2. Open VS Code Settings and configure oaicopilot.baseUrl and oaicopilot.models.
3. Open GitHub Copilot Chat interface.
4. Click the model picker and select "Manage Models...".
5. Choose "OAIProxy" provider.
6. Enter your API key — it will be saved locally.
7. Select the models you want to add to the model picker.

Compatibility note: OAIProxy keeps the existing oaicopilot.* settings keys, so your JSON model configuration stays valid. Because the extension ID changed to lqdflying.oaiproxy, VS Code may require entering API keys once under the new extension.

[!IMPORTANT] Configure a utility model when using OAIProxy/BYOK models in VS Code 1.128+. VS Code leaves BYOK utility routing unconfigured by default, so Agent mode can report: No utility model is configured for 'copilot-utility-small' while the selected main agent model is BYOK. This is a VS Code routing message, not a Kimi K3 or OAIProxy request failure.

Add the recommended option to your User Settings JSON:

"chat.byokUtilityModelDefault": "copilot"

• "copilot" (recommended when GitHub Copilot utility models are available) keeps OAIProxy/Kimi K3 as the main agent model while Copilot handles lightweight background tasks such as intent detection, chat titles, summaries, and commit messages.
• "mainAgent" uses the currently selected OAIProxy/BYOK main model for those utility tasks too. Choose this when Copilot utility models are unavailable; it can send more requests to your provider and may be slower or cost more.

Explicit chat.utilityModel or chat.utilitySmallModel selections take precedence over this default. See the VS Code BYOK utility-model documentation.

Settings Example

"oaicopilot.baseUrl": "https://api-inference.modelscope.cn/v1",
"oaicopilot.models": [
    {
        "id": "Qwen/Qwen3-Coder-480B-A35B-Instruct",
        "owned_by": "modelscope",
        "context_length": 256000,
        "max_tokens": 8192
    }
]

Set max_input_tokens when a provider needs a safety margin below its official context window. OAIProxy advertises that smaller input budget to VS Code and blocks oversized requests before contacting the provider; it does not send max_input_tokens upstream.

Configuration UI

The extension provides a visual configuration interface for managing providers, models, and API keys without editing JSON files manually. Open via the Command Palette (OAIProxy: Open Configuration UI) or click the OAIProxy status bar item.

The Provider Management form lists Azure Foundry first, followed by presets for OpenAI, OpenAI Compatible, TokenRouter, Anthropic, Fireworks, Kimi, DeepSeek, Z.AI GLM, Xiaomi MiMo, and MiniMax. For Azure Foundry, replace the resource-name placeholder with the direct https://<resource>.services.ai.azure.com/openai/v1 endpoint and save its inference key; OAIProxy stores it as oaicopilot.apiKey.azure-foundry and authenticates with api-key. Quick Setup includes concise Azure Foundry cards for Kimi-K2.6 and DeepSeek-V4-Pro. Selecting a preset fills the provider ID, base URL, and API mode; you can still fetch or enter model IDs from the provider's current model list. Example snippets are in examples/openai-responses.jsonc, examples/openai-chat-completions.jsonc, examples/openai-compatible.jsonc, examples/tokenrouter.jsonc, examples/anthropic.jsonc, examples/fireworks.jsonc, examples/zai-glm.jsonc, examples/mimo.jsonc, examples/minimax-openai.jsonc, and examples/minimax-anthropic.jsonc.

OpenAI Compatible (apiMode: "openai-compatible") targets a generic gateway that speaks OpenAI Chat Completions. Replace the placeholder https://YOUR-GATEWAY.example.com/v1 base URL with your gateway's URL (including /v1 when it needs one) and save its key; OAIProxy stores it as oaicopilot.apiKey.openai-compatible and authenticates with Authorization: Bearer. Request shaping follows chathub's OpenAI-compatible provider: only model, messages, stream, stream_options, max_tokens, reasoning_effort, stop, tools, tool_choice and the sampling fields are forwarded, and anything else a gateway needs can be added through extra, which is merged into the body. temperature, top_p, frequency_penalty and presence_penalty are dropped for gpt-6-astra* and gpt-6.1-sol*, matching chathub's pruning for those families. max_completion_tokens is sent as max_tokens, and reasoning_effort is always top-level. When prompt_cache is enabled, OAIProxy derives a stable prompt_cache_key per conversation from the model, reasoning effort, tool selection and tools, every system/developer message, and the first user message. Quick Setup ships one card: GPT-6 Astra with xhigh reasoning, a 258,000-token compatible context budget, and a 128,000-token output allowance. Any other model a gateway serves can be added through Manual Setup; the request shaping above applies to those rows too. A Codex-backed compatible gateway verified on October 30, 2026 no longer serves gpt-5.5 (retired from ChatGPT sign-in on October 14, 2026), so no GPT-5.5 card is shipped. This provider has no public usage/balance endpoint, so the usage check reports an unsupported reason. Responses routing, the Session_id header, image generation, and video pass-through are not part of this provider; use apiMode: "openai-responses" for a Responses-only gateway. See examples/openai-compatible.jsonc.

Quick Setup also includes Grok 4.7 (OAuth) and Grok 4.6 (OAuth). These cards use xAI's subscription Responses route at https://cli-chat-proxy.grok.com/v1; they do not require an API key. In the xAI provider row's Actions, click Sign in, complete the device-code approval in a browser, then add either card. The equivalent commands are OAIProxy: Sign in to xAI / Grok with OAuth and OAIProxy: Sign out of xAI / Grok OAuth. OAIProxy stores the refreshable credential in VS Code SecretStorage and refreshes it before requests. OAuth availability is controlled by xAI account eligibility; xAI API-key models remain available separately through https://api.x.ai/v1.

If Grok OAuth returns 426 Upgrade Required with an outdated Grok CLI version, update OAIProxy. The extension supplies the compatibility version for both chat and usage requests; updating a separately installed Grok CLI does not change it. As of October 2, 2026, the proxy requires at least 1.0.13, which OAIProxy now sends (Grok Build release history).

Quick Setup offers GPT-6.1 Sol for OpenAI API keys and GPT-6.1 Sol (OAuth) for ChatGPT/Codex sign-in. Both use gpt-6.1-sol with apiMode: "openai-responses", enable vision and tools, and default to xhigh reasoning. Thinking Effort offers low, medium, high, xhigh, and max; none and minimal are unsupported. The presets use a 1,050,000-token context window and a 128,000-token output allowance. OpenAI's upstream default is medium; xhigh is OAIProxy's preset choice. See the official model specifications and Codex availability, which depends on your account, rollout, and workspace settings.

The Codex OAuth cards for GPT-6.1 Sol, GPT-6 Astra/Sol/Luna, and GPT-5.6 Sol/Terra/Luna use the separate openai-oauth provider at https://chatgpt.com/backend-api/codex and do not require an OpenAI API key. In the OpenAI OAuth (Codex) provider row's Actions, click Sign in; the device code stays visible in a modal with copy/open actions while you complete approval, then add a Codex card. The equivalent commands are OAIProxy: Sign in to OpenAI / Codex with OAuth and OAIProxy: Sign out of OpenAI / Codex OAuth. Credentials are stored in VS Code SecretStorage and refreshed before requests. OpenAI documents ChatGPT/Codex sign-in eligibility and plan-based usage separately from API-key billing; account availability can change.

GPT-6 Astra Fast (OAuth) is a separate opt-in Quick Setup card that requests OpenAI Fast mode through the Codex OAuth route. OAIProxy keeps service_tier: "fast" as the configured alias and sends the backend-compatible priority value on the wire. Fast mode uses included ChatGPT/Codex subscription usage at 2.5x the Standard rate, or purchased credits and Enterprise pay-as-you-go usage at 2x the Standard rate where available. Standard GPT-6 Astra (OAuth) remains available without the speed premium. Fast mode availability and latency depend on the account, workspace, rollout, and current capacity; GPT-6 Astra Fast mode has no latency SLA. See OpenAI's Codex speed documentation, Responses API Fast mode documentation, and GPT-6 Astra model details.

Codex OAuth prompt caching has special request-affinity and Responses-body requirements. See the OpenAI OAuth cache requirements lesson before changing this provider.

Model Management provides a per-model Test action and a parallel Test all action. Each test sends a small real inference request through the model's saved provider, API mode, headers, and advanced request settings, so it may incur minimal provider usage.

The standalone Provider Usage Check table lists configured supported providers dynamically and reports xAI/Grok weekly credit remaining, Codex OAuth quota windows, credit, token, token-plan, or cost usage. Codex quota uses the saved OAuth credential and OpenClaw-compatible https://chatgpt.com/backend-api/wham/usage bearer endpoint; it is best-effort compatibility support and may become unavailable if OpenAI changes that undocumented route. Fireworks account discovery and month-to-date serverless token checks reuse the normal Fireworks provider key. OpenAI and Anthropic usage/admin keys are stored separately from chat API keys. Anthropic cost uses the Usage and Cost Analytics API (/v1/organizations/analytics/cost) and falls back to the legacy cost report when that export has no rows yet; the analytics export refreshes about every four hours, so the fallback is what shows spend from the last few hours. TokenRouter uses a separate Management Key and reports account credits from its Management API self-wallet endpoint. Azure Foundry inference keys cannot call an official usage/cost endpoint; use Azure Monitor or Cost Management with Azure RBAC. Z.AI and MiMo entries are likewise shown with unavailable reasons when their public docs do not expose API-key usage or balance endpoints.

Provider usage credentials are sent only to validated HTTPS destinations. Built-in adapters use fixed official endpoints, TokenRouter wallet checks are pinned to https://api.tokenrouter.com, and a LiteLLM usage check against any origin other than the default Nube gateway (https://ai.nube.sh) requires explicit confirmation before the privileged Usage Key is sent. LiteLLM usage checks therefore require an HTTPS base URL; chat still works against an HTTP self-hosted gateway, but its usage check reports the HTTPS requirement instead of sending the key.

Configuration export omits the generic API key, provider API keys, and provider usage/admin keys by default, and replaces sensitive custom header values (auth, token, key, secret, credential, or signature in the name) with [REDACTED], so exported files are safe to share. Selecting Export API Keys displays a modal warning and requires confirmation; the resulting file contains those keys and custom header values in cleartext and should be stored securely. Import validates the file before applying configuration, providers, or secrets, and rolls back partial changes if an import step fails.

→ Full Configuration Guide

Multi-API Mode

Supports seven API protocols: azure-foundry (direct Azure Foundry Chat Completions with api-key authentication), openai (Chat Completions), litellm (LiteLLM Proxy Chat Completions), openai-responses (Responses), ollama, anthropic, and gemini. Specify per-model via the apiMode parameter.

Ollama model discovery sends the provider's configured key as a bearer token; the literal placeholder key ollama suppresses the header for unauthenticated local servers. When enable_thinking or thinking.type expresses an explicit intent, OAIProxy sends Ollama's native think field as true or false; leaving both unset preserves the server default.

TokenRouter, Fireworks, Kimi, DeepSeek, Z.AI GLM, and Xiaomi MiMo use the existing openai mode because their hosted APIs are OpenAI-compatible. MiniMax supports both openai and anthropic modes; MiniMax recommends the Anthropic-compatible M3 endpoint for thinking and interleaved-thinking workflows.

→ Full Multi-API Guide

Azure Foundry

Azure Foundry uses its direct /openai/v1 endpoint, not an Azure API Management gateway. Configure the resource URL and inference key in Provider Management, then add the Kimi-K2.6 and/or DeepSeek-V4-Pro Quick Setup cards. The Kimi card defaults to high reasoning; DeepSeek defaults to max. Both cards support tools, and Kimi enables vision.

Azure Foundry performs eligible model prompt caching automatically. OAIProxy deliberately does not send prompt_cache_key, prompt_cache_retention, or thinking for these cards. A cache hit does not prevent the model from receiving the conversation context: Foundry reuses cached prefix computation internally while the model still processes the full logical request. Confirm real hits through cache.usage logs and usage.prompt_tokens_details.cached_tokens; caching is best effort, so identical requests can still report no hit.

Fireworks AI

Fireworks is a first-class provider using https://api.fireworks.ai/inference/v1 and full model IDs such as accounts/fireworks/models/deepseek-v4-pro, accounts/fireworks/models/kimi-k2p7-code, and accounts/fireworks/models/glm-5p2. Quick Setup keeps these Fireworks-hosted cards separate from the original DeepSeek, Kimi, and Z.AI cards.

Fireworks prompt caching is enabled by default upstream. OAIProxy adds a stable user affinity value to improve cache routing and reads cached-token telemetry from OpenAI-compatible usage responses. The Fireworks usage check discovers accounts through the public accounts API and reports month-to-date serverless input/output tokens; the HTTP endpoint does not expose rated cost totals.

TokenRouter

TokenRouter is configured as an OpenAI-compatible provider at https://api.tokenrouter.com/v1. The Quick Setup cards use the exact gateway model IDs deepseek/deepseek-v4-pro-0813, qwen/qwen3.8-max, moonshotai/kimi-k3, and z-ai/glm-5.3; the provider key is stored separately as oaicopilot.apiKey.tokenrouter. TokenRouter credit checks use a separate Management Key stored as oaicopilot.usageApiKey.tokenrouter and call the Management API wallet endpoint.

DeepSeek Flash

The direct DeepSeek Quick Setup card uses deepseek-flash, displayed as DeepSeek Flash, at https://api.deepseek.com in OpenAI Chat Completions mode. It enables vision, tools, and thinking with a 1M context window and max_tokens: 393216 (384K). Thinking Effort offers low, high, and max, with max selected by default; reasoning history is forwarded for tool conversations, and prompt caching is managed by DeepSeek.

This card replaces the legacy Flash names deepseek-v4-flash and deepseek-v4-flash-vision-exp in Quick Setup. Existing saved models remain as configured. See the official model details, Chat Completions API, and vision guide.

Z.AI GLM-5.3 and GLM-5.3-Flash

Quick Setup offers GLM-5.3 (glm-5.3, text-only) and GLM-5.3-Flash (glm-5.3-flash, with vision) through the Z.AI GLM Coding Plan endpoint (https://api.z.ai/api/coding/paas/v4) in OpenAI Chat Completions mode. Both use a 1,000,000-token context window and max_tokens: 131072.

Both presets enable mandatory thinking and tools, offer low, high, and max Thinking Effort with max selected by default, and preserve reasoning history with thinking.clear_thinking: false and include_reasoning_in_request: true. They set temperature: 1, top_p: 0.95, and extra.tool_stream: true for streamed tool calls. Context caching is provider-managed.

These cards replace the direct Z.AI GLM-5.2 Quick Setup card; existing saved models remain as configured. See GLM-5.3, GLM-5.3-Flash, the Coding Plan model guide, and the Chat Completions API. A complete configuration for both models is in examples/zai-glm.jsonc.

MiniMax M3

Use MiniMax-M3 as the model ID. MiniMax's M3 docs list a 1,000,000-token context window, tool use, coding/agentic workflows, adaptive thinking, and native image/video input. OAIProxy supports M3 through the existing MiniMax OpenAI-compatible preset (https://api.minimax.io/v1) and MiniMax Anthropic-compatible preset (https://api.minimax.io/anthropic).

For OpenAI mode, use max_completion_tokens, thinking: { "type": "adaptive" }, and extra.reasoning_split: true to stream thinking separately. For Anthropic mode, use max_tokens and thinking: { "type": "adaptive" }. Set vision: true for M3 so OAIProxy forwards image and supported video LanguageModelDataParts directly instead of routing image inputs through Vision Bridge.

Vision Bridge

Use images with text-only models. OAIProxy automatically describes images via a configured vision-capable model before forwarding as text, with LRU caching (50 entries, ~500KB).

→ Vision Bridge Guide

Multi-Provider Guide

Configure models from multiple providers simultaneously. Use owned_by to group models by provider, with automatic per-provider API key management stored as oaicopilot.apiKey.<provider>.

→ Multi-Provider Guide

Multi-config for the same model

Define multiple configurations for the same model ID via configId (e.g., glm-4.6::thinking and glm-4.6::no-thinking), each with independent settings.

→ Multi-config Guide

Thinking Effort Control

VS Code 1.120+ exposes a per-model Thinking Effort dropdown in the model picker. Enable it with supports_reasoning_effort: true. The direct DeepSeek Flash card offers low/high/max and defaults to max; legacy DeepSeek presets use high/max. Claude Sonnet 4.6 is detected automatically and maps to Anthropic output_config.effort.

→ Thinking Effort Guide

Custom Headers

Specify custom HTTP headers per model provider (API versioning, additional auth, debugging tokens). Merged with default headers on each request.

→ Custom Headers Guide

Custom Request Body Parameters

Use the extra field to inject arbitrary JSON parameters into the API request body for all API modes. Override standard parameters or add provider-specific features.

→ Custom Request Body Guide

Prompt / KV Cache

OAIProxy surfaces provider cache-hit usage in structured logs and applies safe cache request shaping where supported. OpenAI endpoints get a stable prompt_cache_key by default, Fireworks gets a stable user affinity value, and Anthropic-compatible cache_control writes are opt-in via prompt_cache.anthropic.enabled or explicit VS Code cache_control message parts. Azure Foundry, DeepSeek, Xiaomi MiMo, MiniMax OpenAI mode, and Gemini continue to use provider automatic/implicit caching.

OpenAI previous_response_id is kept for conversation state only; OpenAI still bills previous input tokens in the response chain. Use oaicopilot.logLevel: "info" or "debug" and inspect cache.usage log entries to verify actual cache reads/hits.

Model Parameters

Full reference of all 30+ configurable model parameters (id, owned_by, temperature, reasoning_effort, vision, toolCalling, apiMode, etc.).

→ Model Parameters Reference

Logging

OAIProxy always writes extension lifecycle events (install, update, activate) to the VS Code Output panel. Open Output: Show Output and select OAIProxy.

For request/debug logs, add this to VS Code User Settings JSON:

"oaicopilot.logLevel": "debug"

Valid values are off, debug, info, warn, and error. File logs are written to ~/.copilot/oaiproxy/logs/ with daily rotation (logs older than 7 days are automatically cleaned up). Sensitive header values (Authorization, x-api-key, x-goog-api-key) are automatically redacted from log output.

Thanks to

Thanks to all the people who contribute.

• Contributors
• Hugging Face Chat Extension
• VS Code Chat Provider API

Support & License

• Open issues: https://github.com/lqdflying/OAIProxy/issues
• License: MIT.
• Original upstream copyright (c) 2025 Johnny Zhao; OAIProxy changes copyright (c) 2026 lqdflying.