Provider Guides
June 5, 2026 · View on GitHub
This index collects provider-specific guides for configuring VT Code with different LLM backends.
Google Gemini
- Configuration details are covered in the main Getting Started guide.
- Models and constants are defined in
vtcode-core/src/config/constants.rs.
OpenAI GPT
- Official docs:
- Follow the Getting Started guide for API key setup.
- See
vtcode-core/src/config/constants.rsfor the latest supported models. - GPT-5.2 reference: Using GPT-5.2
- VT Code's default OpenAI profile is
gpt-5.4withreasoning_effort = "none"andverbosity = "medium"; raise reasoning only when the task shape justifies the extra latency. - VT Code applies a compact GPT-5.4 prompt contract rather than a verbatim cookbook prompt: compact outputs, low-risk follow-through, dependency-aware tool use, completeness checks, verification, and conditional grounding/citation rules.
- File inputs are supported for native OpenAI Responses API requests through
input_fileparts. - Supported file input fields in VT Code message parts:
file_id,file_data,file_url,filename. file_urlis Responses API only; VT Code rejectsfile_urlwhen a request uses Chat Completions.- VT Code only upgrades local non-image file refs such as
@report.pdfand@"Quarterly Deck.pptx"into structured file attachments for native OpenAI Responses sessions onapi.openai.com. - Remote external document URLs such as
@https://example.com/letter.pdfare also only elevated to structuredfile_urlinputs for native OpenAI Responses sessions. - ChatGPT subscription sessions, OpenAI-compatible endpoints, and other providers keep non-image
@filerefs as plain text plus file-reference metadata so the agent can resolve the path and read the file with tools. - Raw image paths still use the existing multimodal image path flow. Non-image files require explicit
@...references. - Official OpenAI Responses replays now preserve assistant phase metadata for replayed assistant history (
commentaryfor preambles/progress updates,final_answerfor completed answers) when the target GPT model supports it. VT Code does not send this field to Chat Completions, tool/user items, or non-native OpenAI-compatible endpoints. - OpenAI Responses hosted tools currently map through
ToolDefinitionforweb_search,file_search, hostedtool_search, and remotemcp, with hosted config passed through directly on each tool entry. - OpenAI hosted shell mounts are configured through
provider.openai.hosted_shellinvtcode.toml. - Hosted shell skill mounts support both
skill_referenceandinlinebundle entries; VT Code forwards them to OpenAI but does not upload/create hosted skills in this path. - This hosted-shell workflow is separate from VT Code's local
SKILL.mdfilesystem skills. - For large corpora, prefer File Search/Retrieval instead of sending full files inline.
- For spreadsheet-heavy analysis, use Hosted Shell workflows instead of large inline sheet prompts.
Anthropic Claude
- Key management and defaults mirror the Gemini/OpenAI flow in Getting Started.
- Supported model IDs live in
vtcode-core/src/config/constants.rs.
GitHub Copilot
- Guide: GitHub Copilot Managed Auth
- Runtime dependency:
copilotmust be installed and runnable for login/logout - Optional fallback:
ghis only used when VT Code probes an existing GitHub CLI auth session - Commands:
vtcode login copilot,vtcode logout copilot,/login copilot,/logout copilot
OpenRouter Marketplace
- Guide: OpenRouter Integration
- Official docs:
- Default model:
xiaomi/mimo-v2.5-pro(VT Code's default). Xiaomi MiMo V2.5 and V2.5 Pro are also available. - Xiaomi MiMo models:
xiaomi/mimo-v2.5-pro— flagship agentic model, 1M context, reasoning + tool callsxiaomi/mimo-v2.5— omnimodal model, 1M context, reasoning + tool calls
Atlas Cloud
- Guide: Atlas Cloud Integration
- Official docs:
- Integration mode: configure Atlas Cloud through VT Code's
[[custom_providers]]support because the LLM endpoint is OpenAI-compatible. - Base URL:
https://api.atlascloud.ai/v1 - Recommended model: start with
deepseek-ai/deepseek-v4-flash(DeepSeek's latest flash model, 1M context, $0.14/M input tokens).
Xiaomi MiMo
- Provider key:
mimo - Docs: Xiaomi MiMo Platform
- Pricing: Pay-as-you-go · Subscription · Quick Access
- Setup: Set
MIMO_API_KEYor use the MiMo provider in VT Code's configuration - Models:
mimo-v2.5-pro— flagship agentic model, 1M context, deep thinkingmimo-v2.5— omnimodal model (text, image, audio, video), 1M context
Ollama Local & Cloud Models
- Guide: Local Inference Servers (unified
/localcommand) - Setup: Install and run Ollama locally (official install)
- Configuration: Local usage needs no key; set
OLLAMA_API_KEYto access Ollama Cloud - Default model: Any locally available model (e.g.,
llama3:8b,mistral:7b,qwen3:1.7b) - Cloud models: Use IDs like
gpt-oss:120b-cloudwithOLLAMA_BASE_URL=https://ollama.com - Custom Models: Use the
custom-ollamaoption in the model picker to enter any locally or cloud-available model ID - Base URL: Configurable via
OLLAMA_BASE_URLenvironment variable (defaults tohttp://localhost:11434) - Features: Streaming, structured tool calling (including Ollama's web search tools), and thinking traces when
reasoning_effortis enabled
LM Studio Local Server
- Guide: LM Studio Provider Guide
- Server: Enable the OpenAI-compatible Developer server in LM Studio (defaults to
http://localhost:1234/v1) - Environment: Optional
LMSTUDIO_API_KEYwhen auth is enabled; override host/port viaLMSTUDIO_BASE_URL - Default model:
lmstudio-community/meta-llama-3.1-8b-instruct(local inference) - Catalog: Also ships with
lmstudio-community/meta-llama-3-8b-instruct,lmstudio-community/qwen2.5-7b-instruct,lmstudio-community/gemma-2-2b-it,lmstudio-community/gemma-2-9b-it, andlmstudio-community/phi-3.1-mini-4k-instruct, plus any custom GGUF models you expose - Features: Streaming, tool calling, structured output, and reasoning effort passthrough via the shared OpenAI surface
llama.cpp Local Server
- Guide: llama.cpp Provider Guide
- Server: VT Code targets
llama-serverand defaults tohttp://localhost:8080/v1 - Environment:
LLAMACPP_BASE_URLoverrides the endpoint;LLAMACPP_MODEL_PATHenables VT Code-managed startup - Managed startup: VT Code can launch
llama-server -m /path/to/model.gguf --port ...when the endpoint is localhost and a GGUF path is configured - Starter catalog:
gpt-oss-20b,qwen3.6-27b,qwen3.6-35b-a3b,gemma-4-26b-a4b,gemma-4-e4b, andstep-3.5-flash - Features: Streaming, dynamic
/v1/modelsdiscovery, local no-auth defaults, and OpenAI-compatible request handling
Evolink Multi-Model Gateway
- Provider key:
evolink - Official docs: Evolink Docs
- Base URL:
https://direct.evolink.ai/v1 - Auth:
EVOLINK_API_KEYenvironment variable - Setup: Set
EVOLINK_API_KEYfrom Evolink dashboard, then configureprovider = "evolink"invtcode.toml - Models:
evolink/gpt-5.2(default)evolink/gpt-5.5evolink/deepseek-v4-proevolink/deepseek-v4-flashevolink/doubao-seed-2.0-proevolink/gemini-3.1-pro-previewevolink/gemini-3.5-flashevolink/MiniMax-M3evolink/claude-sonnet-4-6evolink/claude-opus-4-8evolink/claude-haiku-4-5-20251001
- Features: OpenAI-compatible gateway exposing many upstream models behind one endpoint. Evolink serves models under bare upstream names (e.g.
gpt-5.2) that collide with VT Code's first-class providers, so curated model IDs are namespaced asevolink/<model>. The provider strips the prefix before sending requests upstream.
Anthropic API Compatibility Server
VT Code provides compatibility with the Anthropic Messages API to help connect existing applications to VT Code, including tools like Claude Code.
- Feature: Anthropic API compatibility server
- Command:
vtcode anthropic-api --port 11434 - Endpoint:
/v1/messages(mirrors Anthropic Messages API) - Environment variables:
ANTHROPIC_AUTH_TOKEN=ollama(required but ignored)ANTHROPIC_BASE_URL=http://localhost:11434ANTHROPIC_API_KEY=ollama(required but ignored)
- Features: Streaming, tool calling, vision support, multi-turn conversations
ℹ Additional provider-specific guides will be added as new integrations land in VT Code.