Skip to content

Model Profiles and Agent Launch

Jump to bottom
Jazzen Chen edited this page May 24, 2026 · 3 revisions

Model Profiles And Agent Launch

Profiles are saved provider routes. Launch combines a profile with an agent, terminal, workspace, and session.

Launch Layout

The current Launch flow is:

  1. Pick an agent from the left rail.
  2. Pick a profile. Direct means native CLI login.
  3. Pick a terminal.
  4. Pick a workspace.
  5. Pick New session or an existing session.
  6. Click Launch.

The profile grid is intentionally compact so many provider routes fit on one screen.

Profile Contents

A profile can contain:

Field Purpose
Provider DeepSeek, DashScope, Kimi, MiniMax, xAI, NVIDIA NIM, Gemini, OpenRouter, Azure, Custom, etc.
Credentials API key or provider-specific auth values
API types OpenAI Responses, OpenAI Chat, Anthropic Messages, Gemini Generate Content
Model list Models shown to the launched agent or model picker
Context metadata Context window and related model metadata when known
Custom model name User-typed model name passed through without validation
Bridge route Agent API shape to provider API shape
Provider options DeepSeek reasoning, proxy opt-in, endpoint overrides, headers

Model Lists, Fake IDs, And Aliases

Some agents need a model list, not only one model. VibeAround can expose a list of models for a profile:

  • built-in known models
  • provider-specific display descriptions
  • context-window metadata when known
  • user-added custom model names
  • compatibility fake IDs that map back to upstream provider models

The agent may display the fake ID or description. VibeAround still maps the selected model back to the upstream model before sending the provider request. If the provider does not recognize a custom model name, the provider error is returned.

The legacy upstreamModel / fakeModelId pair is still supported for compatibility and default-model behavior. It is planned to be replaced in a future 0.7.x profile schema.

API Bridge Routes

VibeAround's local bridge can translate between:

Client shape Provider shape
OpenAI Responses OpenAI Responses, OpenAI Chat, Anthropic Messages, Gemini
OpenAI Chat Completions OpenAI Chat, Anthropic Messages, Gemini
Anthropic Messages Anthropic Messages, OpenAI Chat, OpenAI Responses
Gemini Generate Content Gemini and supported translated routes

Common examples:

  • Codex CLI -> DeepSeek OpenAI Chat Completions
  • Codex CLI -> Kimi Anthropic Messages
  • Claude Code -> DeepSeek OpenAI Chat Completions
  • Claude Code -> MiniMax Anthropic Messages

Claude Suffix Normalization

Claude Code can append internal suffixes such as [1m] to a selected model name. VibeAround strips the suffix before mapping the model to the provider model. This prevents provider errors such as a DeepSeek request receiving deepseek-v4-pro[1m].

Auth Token Conflicts

Claude Code may warn when both a Claude account token and ANTHROPIC_API_KEY are present. That warning comes from Claude Code itself. Use one mode at a time:

  • native Claude account login for Direct mode
  • ANTHROPIC_API_KEY or a VibeAround bridge endpoint for profile/bridge mode

Local API Endpoint

Manual clients can point at:

http://127.0.0.1:12358/va/local-api/<profile-id>/<scope>/<target-api-type>

Use a stable scope per project or client so bridge state, reasoning replay, and tool-call repair do not mix unrelated conversations.

Provider Guidance

Provider Typical use
DeepSeek Codex/Claude via bridge; handles reasoning and suffix normalization
Alibaba DashScope Coding Plan / Token Plan endpoints
Kimi / Moonshot OpenAI-compatible and Anthropic-style routes
MiniMax OpenAI-compatible and Anthropic-style routes
Xiaomi MiMo Regional Token Plan endpoints
xAI / Grok Responses and Chat routes
NVIDIA NIM OpenAI Chat-compatible route
Gemini Gemini native profile
Custom endpoint Bring your own base URL, headers, API type, model list

Clone this wiki locally