Agents and provider profiles

Vibex uses the Agent Client Protocol (ACP) to connect to online Agents. The Agent (who does the work) and the model provider profile (which model, over which route) are managed separately. An Agent can have several provider profiles; one shared Provider can serve several Agents. This page covers the Agent’s own configuration. Shared Providers, route groups, and the local gateway are covered in Providers, route groups, and the local gateway.

​

How the Agents section is organized

Open Config Center → Agents. The Agent list is on the left; the selected Agent’s details are on the right:

​

Add an Agent

The list is grouped by state: enabled, added but disabled, and not yet added. The action on each entry depends on its state: Vibex ships four built-in presets: Claude Code, Codex, ZCode, and OpenCode. The ACP catalog also includes Antigravity, Cline, Codebuddy, GitHub Copilot, Cursor, DeepSeek Harness, Devin, Gemini CLI, Grok Build, Hermes, Kimi Code, Pi, and more.

​

Managed runtimes

• Install downloads and verifies the runtime; it cannot be cancelled midway.
• Check for updates reads the Agent’s release channel and pinned version.
• Upgrade moves only to a compatible release.
• Roll back returns the managed install to the version Vibex pins for that Agent. A rollback is not a hold: the Agent keeps tracking its channel afterwards.
• Uninstall asks for confirmation, removes the managed runtime and cached versions, and keeps session history.

​

Add a custom ACP Agent

Click Add custom ACP Agent at the bottom of the list and fill in: A custom Agent can currently be created but not edited; turning its switch off removes it, while session history is kept.

​

Built-in tools

Each Agent’s built-in tool panels can be toggled from the Agent list. Enabled tools light green:

• Terminal: the Agent may request command execution.
• Browser: the Agent may use the embedded browser tools.
• Computer use: the Agent may use desktop tools (the driver must also be installed and the master switch on).

A tool switch decides whether the Agent receives that tool surface; it does not change existing permission approvals. Turning a tool off does not undo actions already performed.

​

Sign in to an Agent account

Native credentials manages the Agent’s own account, shared with the external CLI—the card says “Agent-native account · shared with the external Agent CLI”.

• The Agent decides which sign-in methods exist: Sign in, Save and sign in, Open sign-in terminal, and others.
• The card offers Refresh auth methods, Verify account, Refresh models, Re-authenticate, and Sign out.
• Re-authenticating affects Agent processes still using the old credential; the UI warns first.
• Sign out affects both the external CLI and Vibex sessions, and shows how many sessions are affected.

Browser-based sign-in (Cursor, Grok, Gemini, Cline, CodeBuddy, Antigravity, and others) may not advertise a logout method; Vibex recognizes the OAuth method as usable instead of treating it as unavailable.

​

Configure a model provider

In Model provider configuration, click Add configuration or edit an existing profile. The editor has three parts:

​

Basics

​

Connection and credentials

The API key field is write-only: masked by default, with an eye icon for a temporary reveal. A saved value is marked Saved. A field you did not change is not rewritten; saving it blank keeps the old value. To remove it, use Clear.

​

Models

The editor has two columns: the available models picker on the left and the selected model’s model settings on the right. Vibex does not ship a model list:

• Fetch models needs an API key and address; it fills the picker without writing configuration.
• Available models: checking a model adds it to the draft; unchecking removes it but keeps declared capabilities, so re-checking restores them. Select all / Enable all handle batches. If the endpoint offers nothing, add a model id under Custom model.
• Model ID / Display name: edited on the right. Selection is tracked by model id, so renaming keeps the right model selected.
• Model wire protocol: each model can override the Provider default or inherit it.
• Reasoning capability / reasoning levels: declare whether the model reasons and which depths it supports.
• Context window / maximum output tokens: declared under Advanced.

If the Provider does not offer a model, the right column shows Add this model. Reasoning levels come from the Provider or Agent declaration; common values are off, minimal, low, medium, high, xhigh, and max. A wrong declaration gives the Agent an option it cannot serve; leave it empty when unsure.

​

Use in a session

Before sending, choose Agent, provider profile, model, and reasoning depth in the composer. If the Agent is bound to a route group, the picker includes the route group’s agent-facing model name. A switch only affects later runs; the existing timeline keeps its original Agent and model identity.

​

Troubleshooting

• The Agent is not in the list: confirm the executable is on PATH, or check the custom command and arguments.
• A model is not selectable: Refresh models, check that the model is enabled, and check whether the Agent is bound to a Provider or route group that includes it.
• Authentication fails: confirm the account state and sign in again; do not paste response headers or tokens into an issue.
• A remote client cannot see the configuration: confirm the device has full_control and check which authoritative runtime you are connected to. Local and remote configuration are separate.
• A managed Agent fails to install: check network access to GitHub and the runtime’s distribution source, then retry.
• The Agent does not receive browser or computer tools: check the Agent’s built-in tool switches, plus the computer-use master switch and driver state.