Skip to main content

Providers

Vanta is model-agnostic. The agent loop only sees the LLMProvider interface, so backends swap by changing VANTA_PROVIDER + VANTA_MODEL — no code changes.

Supported backends​

VANTA_PROVIDERNotes
openaiOpenAI API; GPT-5 and o-series models get reasoning_effort from the effort level
ollamaLocal models, no API key — the default (qwen2.5:14b)
anthropicAnthropic Claude models; thinking-capable models use extended thinking on high/max effort
geminiGoogle, via the OpenAI-compatible adapter (baseURL swap)
openrouterMany models behind one key, via the OpenAI adapter
azureAzure OpenAI / AI Foundry (deployment + api-version)
customAny OpenAI-compatible endpoint — point VANTA_OPENAI_BASE_URL at it

Plus subscription-OAuth backends (codex for ChatGPT, claude-code for a Claude Pro/Max plan) and named OpenAI-compatible shortcuts (deepseek, xai, groq, mistral, together, fireworks, cerebras, nvidia, moonshot, minimax, perplexity, tokenrouter, openrouter, and more — each just a baseURL + key swap).

Set it interactively with vanta setup, or edit vanta-ts/.env.

Current OpenAI models​

Vanta's OpenAI API and ChatGPT-subscription Codex pickers include the current GPT-5.6 family: gpt-5.6-sol (the default for difficult coding and research), gpt-5.6-terra (balanced everyday work), and gpt-5.6-luna (fast, repeatable work). The OpenAI API picker also includes current API-key aliases and dated snapshots such as gpt-5.5-2026-04-23, gpt-5.4-2026-03-05, gpt-5.2-2025-12-11, gpt-5.1-2025-11-13, gpt-5-pro-2025-10-06, gpt-5-chat-latest, o1-pro, and computer-use-preview. The Codex subscription picker keeps Codex-agent models under the OAuth-backed codex provider, including gpt-5.3-codex, gpt-5.3-codex-spark, gpt-5.2-codex, gpt-5.1-codex-max, gpt-5.1-codex, gpt-5.1-codex-mini, gpt-5-codex, and gpt-5-codex-mini.

The picker is curated rather than an access-control boundary: type any model ID your account can use when it is not shown.

Session vs default model​

Model switches are session-scoped by default. They hot-swap the current conversation and are saved with that session, but do not mutate process.env or vanta-ts/.env.

/model ollama qwen2.5:14b             # this session (default)
/model openai gpt-4o --session        # explicit session scope
/model gemini gemini-2.5-pro --global # set current + future default
/model --global                       # set the current session model as default

The TUI and desktop pickers use session scope for their primary model action and expose Set as default separately. Two concurrent sessions can therefore use different providers; resuming either restores its saved route. Sessions created before this metadata existed adopt the configured default and write explicit model metadata on their next save.

Routers reach every model​

A router (tokenrouter, openrouter) is one key that proxies many models, so you're never pinned to one. VANTA_MODEL accepts any id the router serves — and vanta setup free-types the model for a router instead of showing a fixed list. (VANTA_MODEL has always accepted any string; routers just make that the default UX.)

Add any provider — no code (~/.vanta/providers.json)​

You don't need to touch the source to add a backend. Declare it in ~/.vanta/providers.json, the same way ~/.vanta/agents.json adds agent CLIs and ~/.vanta/mcp.json adds MCP servers:

{ "providers": {
  "nvidia-free": { "baseURL": "https://integrate.api.nvidia.com/v1", "keyEnv": "NVIDIA_API_KEY", "model": "deepseek-ai/deepseek-r1" },
  "tokenrouter": { "baseURL": "https://api.tokenrouter.com/v1", "keyEnv": "TOKENROUTER_API_KEY", "model": "MiniMax-M3", "router": true },
  "lan-llm":     { "baseURL": "http://192.168.1.5:8000/v1", "model": "qwen2.5:32b" }
}}

Then VANTA_PROVIDER=nvidia-free (+ optional VANTA_MODEL). Rules:

  • The secret never lives in this file — only the env-var name (keyEnv); the actual key stays in vanta-ts/.env.
  • Omit keyEnv for keyless local endpoints.
  • A user entry overrides a built-in of the same id — so you can also fix a built-in's URL or default model yourself without waiting on a release.
  • router: true makes the setup wizard free-type the model.

Template: vanta-ts/providers.json.example.

Effort levels​

VANTA_EFFORT_LEVEL (or /effort <level> live) is low | medium | high | xhigh | max:

  • OpenAI GPT-5 and o-series — maps to reasoning_effort for low/high/xhigh/max when the selected model supports it.
  • Anthropic — high/max enable extended thinking (8k / 32k budget tokens); low caps max_tokens.

Multimodal​

User messages can carry images (attach via /image, /paste, or drag-drop). Every provider maps them natively. Image-understanding tools resolve a dedicated vision model via VANTA_VISION_MODEL when set, so a text-only main model still has working vision.

Task routing (optional)​

Set VANTA_MODEL_CHEAP / VANTA_MODEL_EXPENSIVE and Vanta classifies each task and routes it to the cheaper or stronger model. Unset = no routing.

Per-call route usage​

Every completed agent-loop provider call is appended to <project>/.vanta/route-usage-ledger.jsonl. Each versioned row records the session and surface, actual serving provider/model, normalized base route, fallback depth, billing mode/status, input/output/cache/reasoning tokens, one API call, and cost when known.

Model catalog updates​

The desktop model picker refreshes from Vanta's published model catalog and caches it for six hours. If the network is unavailable, it uses the last valid local catalog and then the bundled catalog. A listed model still requires that your selected provider account is entitled to run it; Vanta reports provider errors rather than treating a picker entry as an entitlement guarantee.

This is call-level attribution, so a fallback is charged to the route that served it and a mid-session model switch starts a new route row. Local and subscription-included calls remain visible at $0; unknown-cost calls remain visible as ~?.

/usage breakdown   # route totals when route rows exist; legacy spend view otherwise
/dashboard         # current model plus accumulated serving routes

The older spend-ledger.jsonl remains the turn-level budget compatibility source. Views never add both ledgers together, so existing sessions remain readable without double counting.

Adding a provider​

Implement LLMProvider (complete / modelId / contextWindow) and add a branch in providers/index.ts. The loop stays unchanged — see Extending Vanta.