Agents

Settings → Agents governs the installation, enabling, authentication, models, environment variables and native configuration of coding agents.

• Title: Agent Management
• Subtitle: Manage Agent SDK connection, enabled state, environment variables, config management and version preflight info in one place.

Only four agents appear on this screen

What you can see in the interface is these four, in this order: Codex CLI, Claude Code, Grok, DeepSeek Harness.

Layout

At the top is a row of draggable agent pills. The pill itself is a single selector — an icon with a small status dot, greyed out when the agent is not enabled, and no switch on it. Drag them to change the order they appear in inside the workspace.

The enable switch is in the title row of the detail card below. That row, in full: icon, name, distribution badge, ACP adapter badge (only for agents that go through an adapter), Custom badge, enable switch, and a one-line description.

Sections every agent has

Version preflight

Shows the number of preflight items, with two buttons: Refresh check and Diagnose (which opens a dialog listing the result of each item). Before it has ever run it shows Unchecked.

Installation and version

Version status has a dozen or so values; the common ones are: not installed / local only / upgrade available / already latest / a newer unreviewed release / current platform does not support this agent / uv runtime not ready, plus edge cases such as "Local version is not comparable" and "No recommended version is available".

The matching actions: Install / Upgrade / Uninstall / Custom install / Upgrade to unreviewed latest. Installation streams a live log.

Environment Variables

The environment variables passed to that agent, set per agent; press save when you are done.

Let the agent handle files and commands

A single switch (what it actually writes is CROLI_ACP_HOST_TOOLS=agent|default). With it on, file reads and writes and command execution all go back to the CLI itself, and this agent no longer gets the delegation tools — the Collaboration screen lists the names of the agents affected.

Config Management / Native JSON Config

Edit the CLI's own configuration file directly (config.toml, settings.json, auth.json and so on, depending on the CLI). This is an escape hatch: whatever Croli's interface does not cover, you write by hand here.

Authentication method

One shared set of options:

Pick Model Provider with none configured and it points you at that screen to add one. Only Claude Code and Codex have this set of tiers; for the other two agents see their own sections below.

What is special about each of the four

• Claude Code — authentication method; default models for Main Model, Reasoning (thinking), Haiku, Sonnet, Opus and Fable; custom model ID and name; whether to send the attribution identifier header; disabling non-essential traffic.
• Codex CLI — authentication method plus ChatGPT sign-in (a device code flow); provider and model name; switches for WebSocket, Skills, Allow questions in Default mode, and Fast; a whole Sandbox & approvals group — approval policy (on-request / untrusted / never / granular, with 5 further items under granular), sandbox mode (read-only / workspace write / full access), extra writable folders, allow network access, and excluding TMPDIR and /tmp; native auth.json and config.toml editors.
• Grok — permission mode (Ask every time / Auto-approve edits / Smart auto-approve / Always approve); reasoning effort low/medium/high/max; authentication (XAI_API_KEY or a custom endpoint, or grok login); a bring-your-own model (model ID, base URL, API backend Responses / Chat Completions / Messages(Anthropic), API key, context window, whether it is the default for new sessions); auto-compact threshold; native config.toml.
• DeepSeek Harness — API endpoint plus API key.

Multi-Agent Collaboration

At the very bottom of the page is the delegation panel (enable delegation, maximum delegation depth, completed-result cache, agent defaults). The same panel also hangs on the Collaboration screen, and the detailed explanation is there.

Worth knowing

• Where an agent's skill files go differs from CLI to CLI, see Plugins and skills. More agents can hold skills than the four on this screen, so skill-related pages filter by which agents can hold skills — a different list.
• The enable switch, the ordering and the environment variables are all stored on the backend.
• The ACP adapter badge means: this CLI does not speak Agent Client Protocol itself, so Croli translates for it with an adapter.