Supported AI agents

init writes CLAUDE.md (which points to AGENTS.md), the slash commands under .claude/commands/, the skills under .claude/skills/, project settings in .claude/settings.json and the MCP servers in .mcp.json. Start with /camel-start. During /camel-execute, independent tasks of the same wave are dispatched to parallel subagents with isolated context windows.

Use --ai bob2 for both Bob Shell and Bob IDE; the generated assets are verified with Bob Shell 2.0.2 and Bob IDE 2.1.0. init writes .bob/custom_modes.yaml, rules, the shared skills under .bob/skills/, scoped agent presets under .bob/agents/, the role personas under .bob/personas/ and the MCP servers in .bob/mcp.json.

In Bob Shell, start from the project root, open /skills or type $camel-, and select one of the nine public skills: camel-start, camel-brainstorm, camel-migrate, camel-plan, camel-execute, camel-validate, camel-ship, camel-knowledge and camel-debug. Enter the request after the skill name:

Bob Shell’s native skill picker uses $camel-*; its / picker holds built-in commands and MCP prompts, so a /camel-* name in this documentation corresponds to a $camel-* skill there. Bob IDE exposes the same skills as /camel-*. Legacy command stubs stay installed for compatibility. The four internal helper skills are hidden.

If a public skill does not appear in the picker, run camel-kit doctor: it reports each affected SKILL.md as FAIL with regeneration instructions. Regenerate from the project root with camel-kit init --here --ai bob2 --force (after committing or backing up customized generated files) and restart Bob.

Install Antigravity following Google’s getting-started guide, then:

init writes AGENTS.md, the skills under .agents/skills/, three custom agents under .agents/agents/ (camel-worker with file and sandboxed command tools, camel-reviewer with read and search tools plus the inherited MCP access, camel-ship-worker with read and search tools for Ship stages), the role personas under .agents/camel-kit-personas/ and the MCP servers in .agents/mcp_config.json (two by default; three with --with-citrus-mcp). Open the project root in Antigravity and invoke /camel-start in the CLI, or ask it to use the camel-start skill by name; /skills lists the native skills and /mcp the connections.

The primary conversation owns questions, approvals, orchestration and report writes, and dispatches independent tasks with the native invoke_subagent; the children inherit the workspace permissions and MCP servers. Tool calls stay subject to Antigravity’s permissions: Camel Kit generates no auto-approval policy. Reinitialization preserves unrelated MCP servers, settings and comments in mcp_config.json, which must contain a JSON object ({} when empty).

init writes AGENTS.md, the skills under .agents/skills/, the custom agent roles under .codex/agents/ (planner, implementer, tester, validator, migrator, catalog researcher and security reviewer; the development snapshot adds a Ship worker) and the MCP servers in .codex/config.toml (two by default; three with --with-citrus-mcp). It does not create .codex/commands/: Codex discovers the skills directly, so you start with $camel-start after trusting the repository, and inspect skills with /skills and servers with /mcp.

Codex loads the repository .codex/ configuration only for trusted projects. Review the repository before trusting it; until then the custom agents and the MCP servers are not active. The setup stays repository-scoped: Camel Kit does not edit ~/.codex/config.toml, credentials or login state, generates no hooks, and does not loosen the sandbox or the approval policy. Keep the sandbox enabled and grant only the narrow approval needed when a command is blocked: repository trust lets the project configuration load, it is not a reason to approve every command. Each generated MCP server table carries an exact tool allowlist and default_tools_approval_mode = "prompt". If .codex/config.toml already exists with valid unrelated settings they are preserved; invalid TOML or a conflicting server table makes init fail without changing the file.

camel-kit doctor validates the TOML, the server tables and their allowlists, the prompt approval defaults, the skills, the required custom-agent fields and the read-only sandbox of the research and security agents.

init writes .github/copilot-instructions.md, the skills under .github/skills/, the custom agents under .github/agents/, the MCP servers in .github/mcp.json and a safety hook in .github/hooks/camel-kit-safety.json that denies destructive or secret-sensitive shell commands while leaving the normal permission prompts in place. Start by asking Copilot to "Use the /camel-start skill"; /skills list shows the project skills.

init writes AGENTS.md, the skills under .pi/skills/, prompt templates under .pi/prompts/, .mcp.json for pi-mcp-adapter, and a guard extension (.pi/extensions/camel-kit-guard.ts with its JSON policy). In Pi, trust the project with /trust, then run /skill:camel-start. Headless checks must use pi -a.

Pi has no native subagents: the pipeline runs in the main session and records the missing isolation. For isolated read-only research, start a separate Pi session with constrained tools, for example pi --tools read,grep,find,ls, and bring the findings back.

init writes QWEN.md, the commands under .qwen/commands/, the skills under .qwen/skills/, five bounded leaf agents under .qwen/agents/ (implementer, reviewer, tester, validator, ship-worker), the personas under .qwen/camel-kit-personas/ and the MCP servers in .qwen/settings.json. Start with /camel-start. Interviews, approvals and phase handoffs stay in the primary session; only bounded work goes to the leaf agents. Eligible sessions relay Ship controller stages to the read-only camel-ship-worker named agent via --backend qwen-native; the worker returns structured results and artifact proposals without editing files or calling MCP tools.

init writes AGENTS.md, the commands under .opencode/commands/, the skills under .opencode/skills/, the agent definitions under .opencode/agents/, the personas under .opencode/camel-kit-personas/ and the MCP servers in opencode.json. Start with /camel-start. When a project configuration already exists (opencode.json, opencode.jsonc, or the same files under .opencode/), init preserves unrelated settings, permissions and MCP servers and writes its own entries to the highest-precedence file; invalid or conflicting configuration fails before any file is changed. The generated executor can delegate to allowlisted child agents one level deep; you can also address them directly with the @agent-name syntax.

Camel Kit 0.4.0 replaced the gemini target with Google Antigravity and removed IBM Bob v1. Reinitialize an existing project from its root with the replacement target, then check it:

Before reinitializing, .agents/mcp_config.json must be valid JSON without duplicate keys (doctor reports duplicates as failures), and AGENTS.md, GEMINI.md and everything under .agents/ must be real files or directories, so replace symbolic links first; these checks run before any project file is changed. --force replaces the generated AGENTS.md, skills and configuration, so preserve custom additions to those files first; custom project rules can live in GEMINI.md, which Antigravity also reads. Antigravity reinitialization removes the four old Camel Kit @.gemini/…​ imports from GEMINI.md and keeps custom text, imports and other .gemini/ files. Move custom MCP servers from .gemini/settings.json to .agents/mcp_config.json using Google’s MCP schema, and custom skills from .gemini/skills/<name>/ to .agents/skills/<name>/. Bob 2 reinitialization replaces the generated gate skills and modes with the shared pipeline and native subagents, removing only the known obsolete Bob v1 mode-rule files.