Skip to content
Browse documentation

Connections & models

Choose and configure a model connection

Connect ChatGPT, APIs, compatible endpoints, local models, or supported CLIs.

Connections define how Vopit obtains a model. The effective model may come from the workspace default or an agent override. Account subscriptions, API billing, local inference, and CLI authentication are separate concepts; do not assume that signing in to one grants access to another.

Settings → Connections → AI connections groups connections into Accounts & CLI (ChatGPT, Codex CLI, Claude Code, Mistral Vibe, Mammouth Code, Antigravity CLI), API connections (OpenAI API, Claude API, Mistral API, Mammouth AI, Compatible API) and Local models (Ollama). A fresh project uses Ollama as its default connection, with no model selected.

Available connection types

ConnectionWhat Vopit configuresImportant distinction
ChatGPTBrowser account connection, account/workspace selection, available models, default and fast model.Model eligibility comes from the connected account and can change.
OpenAI API, Claude API, Mistral API or Mammouth AIAPI key file, model discovery, default and fast model, optional pricing. Claude API and Mammouth AI offer Get API key, which opens the provider's key page.API billing is separate from a chat subscription.
Compatible API (API connections) or Ollama (Local models)A HTTPS endpoint, or HTTP only on localhost; model catalogue, selected model and fast model.The endpoint must expose the supported API shape and a tool-capable model.
Codex CLIInstalled executable, sign-in, model catalogue, reasoning effort, and native sessions.Codex controls its internal turns and context management.
Claude CodeInstalled CLI, sign-in, model catalogue of Claude aliases and pinned IDs, reasoning effort at connection and agent level, and a response probe.Runs native tools in the selected project with resumable sessions. An agent can also be granted MCP servers from your own Claude Code configuration.
Mistral Vibe / Mammouth CodeInstalled native executable, CLI configuration, optional model or alias, and a response probe.Each CLI runs its native tools in the selected project with resumable sessions.

Granted Vopit app actions complement every native CLI; see Use native coding CLIs.

Configure an API or compatible endpoint

  1. Open Settings → Connections, select API connections or Local models, and choose the provider. Use Import key file (a .txt or .key file containing only the key) or Replace key file; never paste a key into a request, agent instruction, or repository file.
  2. For a compatible endpoint, use HTTPS. HTTP is accepted only on localhost, 127.0.0.1, or ::1. API paths such as /v1 are supported; URL credentials, queries, and fragments are rejected.
  3. Choose Save & load models, then select an available default model. A failed catalogue response usually points to the endpoint, credentials, or network; a successful catalogue is not an inference check.
  4. Set an agent override only where that agent needs a different provider or model.

Fast model and request budgets

  • Each connection has a Fast model. The Ultra-fast and Balanced speed modes run their quick stages and meeting moderation on it. Leave it empty to use a known lighter model for that connection (for example Sonnet for Claude Code) or else the connection model. Big brain does not use it.
  • Request budgets and model capabilities sets the model catalogue timeout (default 30 s) and connection test timeout (default 60 s). Connections other than CLIs also set the response token budget (default 8,192 tokens) and temperature and reasoning support overrides. Restore connection defaults resets these values.
  • For OpenAI API and Claude API on their official endpoints only, Vopit can use an OPENAI_API_KEY or ANTHROPIC_API_KEY environment variable when no key is stored. Custom endpoints never read these variables.
  • Native CLI processes start without inherited API-key, token, secret, password or credential environment variables, so a CLI uses its own sign-in and does not bill an unrelated key by accident.

Read the effective engine and loading states

Settings → Connections has two parts: AI connections for models and Infrastructure for SSH and database targets. How your agents run resolves each enabled or disabled definition using its connection override or the project default, and shows the actual connection and tools of each agent. Codex CLI, Claude Code, Mistral Vibe, Mammouth Code and Antigravity CLI delegate project reads, edits and commands to their native tools and sessions; Vopit still coordinates the team. ChatGPT and HTTP APIs, including Mistral and Mammouth APIs, supply responses to the Vopit loop. Git, infrastructure, board and Crank actions, the code graph and project memory remain Vopit gateway actions even during a native CLI stage, each limited to the agent's grants.

Catalogues load asynchronously and use a shared 60-second in-memory cache. Refresh models requests fresh discovery. Loading does not save the draft or change the project default by itself. Keep using other controls while a catalogue or probe runs; inspect its own status/error. Settings cannot bypass active-work safeguards for changing a team, project or account.

All documentation