Skip to content

supported_agents.md

Latest commit

 

History

History
199 lines (130 loc) · 8.52 KB

supported_agents.md

File metadata and controls

199 lines (130 loc) · 8.52 KB
Raw
title Supported Agents
description Which AI agents work with Faber and how they differ
icon bot
order 3

Supported Agents

Faber orchestrates multiple AI coding agents through a unified interface. Each agent runs in its own PTY terminal session with full MCP (Model Context Protocol) integration for progress reporting.

Agent Overview

Agent CLI Command Default Model System Prompt MCP Config
Claude Code claude sonnet CLI flag .mcp.json
Codex CLI codex gpt-5.3-codex Instruction file .codex/mcp.json
Copilot CLI copilot (Copilot default) Instruction file .copilot/mcp-config.json
Cursor Agent agent / cursor-agent claude-4-opus Instruction file .cursor/mcp.json
Gemini CLI gemini gemini-2.5-pro Instruction file .gemini/settings.json
OpenCode opencode (user-specified) Instruction file opencode.json

All agents are auto-detected from your system PATH. You can see which agents are available in the session launcher.

Claude Code

Claude Code is Anthropic's official CLI agent.

Supported models: opus, sonnet, haiku, sonnet[1m]

How it works:

  • System prompt is passed via the --system-prompt CLI flag
  • MCP tool documentation is also written to a CLAUDE.md file in the working directory (using <!-- Faber:MCP --> markers to preserve your own content)
  • MCP config is written to .mcp.json in the working directory
  • Model is selected with the --model flag

Codex CLI

Codex CLI is OpenAI's open-source coding agent.

Supported models: gpt-5.3-codex, gpt-5.2-codex, gpt-5.1-codex-max, gpt-5.2, gpt-5.1-codex-mini

How it works:

  • System prompt is written to an AGENTS.md file in the working directory (Codex reads this automatically)
  • MCP config is written to .codex/mcp.json
  • Model is selected with the --model flag

Copilot CLI

Copilot CLI is GitHub's agentic coding assistant for the terminal.

Supported models: claude-sonnet-4-5, claude-opus-4-6, gpt-5.3-codex, gemini-3-pro

How it works:

  • System prompt is written to an AGENTS.md file in the working directory (Copilot reads this automatically)
  • MCP config is written to .copilot/mcp-config.json
  • Model is selected with the --model=MODEL flag
  • Uses -i (interactive with initial prompt) to keep the session alive and ensure MCP connectivity
  • Supports --autopilot for autonomous continuation and --allow-all-tools to skip tool confirmations

Cursor Agent

Cursor Agent is the CLI version of Cursor's AI coding assistant.

Supported models: claude-4-opus, claude-4.5-sonnet, gpt-5, gpt-5.1, gemini-3-pro, gemini-3-flash

How it works:

  • System prompt is written to an AGENTS.md file in the working directory
  • MCP config is written to .cursor/mcp.json
  • Model is selected with the --model flag
  • Faber checks for both agent and cursor-agent binary names during detection

Gemini CLI

Gemini CLI is Google's AI coding agent.

Supported models: gemini-2.5-pro, gemini-2.5-flash, gemini-3-pro, gemini-3-flash

How it works:

  • System prompt is written to a GEMINI.md file in the working directory
  • MCP config is written to .gemini/settings.json
  • Model is selected with the --model flag

OpenCode

OpenCode is an open-source terminal AI assistant.

Supported models: None built-in — you specify the model directly in the agent config.

How it works:

  • System prompt is written to an AGENTS.md file in the working directory (Faber uses <!-- Faber:MCP --> markers so your own edits are preserved)
  • User prompt (initial task message) is passed via the --prompt CLI flag
  • MCP config is written to opencode.json in the working directory
  • Model is selected with the --model flag

How System Prompts Work

Faber composes a system prompt for each session that includes your project's IDE instructions (from .agents/prompts/prompt.md) and MCP tool documentation. The content is tailored to the session mode — task sessions get completion workflow instructions, research sessions get research-specific guidance, and vibe/chat sessions get a lighter set.

How this prompt reaches the agent depends on the agent:

  • CLI flag agents (Claude Code): The prompt is passed directly as a command-line argument (--system-prompt). Faber also writes MCP documentation to CLAUDE.md using marker comments.
  • Instruction file agents (Codex, Copilot, Cursor, Gemini, OpenCode): The prompt is written to the agent's instruction file in the working directory (AGENTS.md or GEMINI.md). Faber uses <!-- Faber:MCP --> markers so your own edits to these files are preserved.

When a session ends, Faber cleans up the instruction file — removing only the Faber-managed section.

MCP Integration

All agents receive MCP (Model Context Protocol) configuration that connects them back to Faber. This lets agents report their progress, file changes, and completion status in real time.

The MCP config is written to the agent's expected config location before the session starts. Faber merges its config with any existing user-defined MCP servers — only the "faber" entry is managed. When the session ends, only the Faber entry is removed.

Available MCP Tools

Agents can call these tools to communicate with Faber. The tools available depend on the session mode — agents only see tools relevant to their session type.

Status & Progress (all sessions)

Tool Purpose
report_status Set working status, message, and activity type. Call first when starting work.
report_progress Report step N of M with description. Drives the progress bar in the UI.
report_files_changed List files that were created, modified, or deleted
report_error Report a hard blocker. The agent should stop and wait after calling this.
report_waiting Signal that user input is needed. The session pauses until the user responds.

Task Management (all sessions)

Tool Purpose
get_task Fetch task metadata and full markdown body
update_task Update task metadata (status, priority, labels, etc.)
update_task_plan Update the implementation plan section of a task file
create_task Create a new task in the current project
list_tasks List tasks in the project with optional status/label filters

Completion (session-mode-specific)

Tool Available in Purpose
report_complete Task, Queue Signal that the task is fully done. Moves the task to In Review. In queue mode, auto-launches the next task.
report_researched Research Signal that research is complete. The user is prompted to continue to implementation. May move the task from Backlog to Ready.

Breakdown, Vibe, and Chat sessions have no completion tool — the user drives the lifecycle.

Agent Configuration

Per-Project Defaults

In Settings > Project, you can set a default agent and model. All new sessions will use this agent unless overridden at launch time.

Per-Task Overrides

Each task can specify an agent and model in its frontmatter. This takes priority over the project default.

Model Resolution Order

When launching a session, the model is resolved in this order (highest priority first):

  1. Manual override in the session launcher
  2. Agent config (per-task or per-project)
  3. Task-level model setting
  4. Project default model
  5. Agent's built-in default model

Requirements

Each agent must be installed separately and available in your system PATH. See each provider's documentation for installation instructions:

You can verify which agents are detected from the session launcher — only installed agents appear as options.