Install an Agent runtime

This page is the entrance to the installation, and it is matched with:

• Daemons and Runtimes — How detection works
• AI Programming Tool Matrix — Differences in capabilities of each tool (session continuation, MCP, model selection)

The Multica server never touches your API key, nor the tool itself. The following operations - installation, login, model access - all happen on your local machine. Problems are almost always local.

1. **Multica daemon is running. ** After installing Multica CLI, run multica daemon start; or use Multica desktop, which will automatically start the daemon process when it starts. If the daemon is not up, no one will scan the tool.
2. **The executable file of the tool is on PATH. ** The daemon invokes the tool with the name shell out (see the command name on the daemon scan line in each section). which <name> cannot be found in the terminal, nor can the daemon. After installation, open a new terminal (or restart the daemon) to allow the new PATH to take effect.

After installing a tool, restart the daemon:

multica daemon restart

For desktop, just restart the app. The daemon scans PATH only once at startup.

Claude Code (Anthropic)

The most complete integration. Session continuation is easy to use, MCP is easy to use, and it will consume the mcp_config field in the agent configuration (see Matrix).

Daemon scan	claude
Installation	See official instructions claude.com/claude-code. The common installation method is the npm package @anthropic-ai/claude-code (requires Node.js 18+).
Authentication	Run claude once and follow the login process in the CLI; or set ANTHROPIC_API_KEY.
Remarks	First choice for new users.

Codex (OpenAI)

JSON-RPC 2.0 transmission, the approval granularity is finer. MCP configuration will be written to $CODEX_HOME/config.toml of a single task. Session resume available - Multica resumes through thread/resume of Codex app-server; it will fall back to a new thread when the thread expires or does not exist.

Daemon scan	codex
Installation	See official instructions github.com/openai/codex.The common installation method is the npm package @openai/codex。
Authentication	codex login (browser login), or OPENAI_API_KEY.

Cursor（Anysphere）

CLI counterpart of the Cursor editor. Session resume available - The current Cursor Agent will return session_id in the stream-json event, and Multica will pass it back with --resume <id> the next time it is run.

Daemon scanning	cursor-agent
Installation	First install Cursor Editor, and then install CLI according to the instructions of docs.cursor.com. The executable file is called cursor-agent, not cursor.
Authentication	Log in in the Cursor editor, CLI reuses the same session.

GitHub Copilot

The model is based on the entitlement of your GitHub account - the tool itself does not select the model, GitHub decides which model you get.

Daemon scan	copilot
Installation	See GitHub's CLI documentation github.com/github/copilot-cli.
Authentication	Log in via GitHub browser in CLI.
Note	The login account must have a valid GitHub Copilot subscription.

OpenCode (SST)Open source CLI agent. Available models will be dynamically discovered from your own configuration files - suitable for users who want to control the model list themselves. The mcp_config field in the agent configuration will be consumed through OPENCODE_CONFIG_CONTENT.

Daemon scan	opencode
Installation	See the official guide opencode.ai or the repository github.com/sst/opencode. Usually a script or npm package is installed.
Certification	Match your own model provider (Anthropic, OpenAI, etc.) according to OpenCode's documentation.

Kiro CLI (Amazon)

ACP-over-stdio transport. Session continuation works through ACP session/load; MCP configuration is passed in through ACP mcpServers; skills are copied to .kiro/skills/.

Daemon scan	kiro-cli
Installation	See Kiro's documentation kiro.dev. The executable is called kiro-cli, not kiro.
Certification	Based on AWS account, follow Kiro's own guidance.

Kimi (Moonshot)

ACP protocol agent, mainly for the Chinese market. MCP configuration is passed in through ACP mcpServers; skills are placed in .kimi/skills/ (native discovery path).

Daemon scan	kimi
Installation	See the official instructions github.com/MoonshotAI/kimi-cli.
Authentication	Moonshot API key, configured according to the manufacturer's documentation.

Hermes (Nous Research)

ACP protocol agent (shares transport layer with Kimi). Session continuation is available, MCP configuration is passed in via ACP mcpServers. Skill injection uses the common fallback path .agent_context/skills/ - verify that the skills are actually loaded before using it.

Daemon scan	hermes
Installation	See Nous Research's repository github.com/NousResearch to get the latest CLI.
Certifications	Per manufacturer documentation.

OpenClaw

Open source CLI agent orchestrator. MCP configuration is written through Multica's single-task configuration wrapper. The model is tied to the agent layer (openclaw agents add --model) - cannot be overridden by task, and --model / --system-prompt cannot be passed from Multica.

| | | |---|---|| Daemon scanning | openclaw | | Installation | See the project github.com/openclaw-org/openclaw (community maintenance). | | Certification | Match the underlying model provider according to OpenClaw's documentation. |

Pi（Inflection AI）

Minimalist style. The method of session resume is different - resume id is the session file path on the disk, not a string id.

Daemon scan	pi
Installation	See Inflection's CLI documentation pi.ai.
Certifications	Per manufacturer documentation.

CodeBuddy (Tencent)

A CLI agent compatible with Claude Code. Multica drives it using the same stream-json protocol as Claude Code: session resume is available via --resume, MCP configuration is passed in via --mcp-config, and skills are placed in .claude/skills/. The model is dynamically discovered.

Daemon scan	codebuddy
Installation	See the official CLI documentation codebuddy.ai/cli.
Certifications	Per manufacturer documentation.

Qoder (Alibaba)

An agentic programming CLI that uses the ACP protocol over stdio (shares the transport layer with Hermes, Kimi, and Kiro CLIs). Session resumption works through ACP session/resume, MCP configuration is passed in through ACP mcpServers, the model is dynamic discovery, and skills are copied to .qoder/skills/.

Daemon scan	qodecli
Installation	See the official CLI documentation qoder.com/cli.
Certifications	Per manufacturer documentation.

Antigravity (Google)Google's Antigravity CLI (agy). With the Google Antigravity service, the Gemini series model is used by default. Multica uses agy -p to start it, which is a one-time non-interactive mode suitable for daemon background tasks; currently Antigravity CLI can still execute tools in this mode, while agy -i requires a TTY connection and is not suitable for daemon drivers. Session continuation works via --conversation <id> - the daemon grabs the conversation UUID from the CLI log file. Model selection is saved in the Antigravity CLI's own settings - the tool's "Model" selection is disabled in Multica. Skill files are written to .agents/skills/ (the CLI follows the workspace layout of the Gemini CLI - see Antigravity documentation).

Daemon scan	agy
Installation	See the official instructions antigravity.google/docs/cli-overview. The CLI is precompiled, just run agy install to configure PATH and shell aliases.
Authentication	Run agy interactively to go through the Google account login process; or log in through the Antigravity desktop - the CLI will reuse the credentials written into the keyring by the GUI.
Remarks	CLI's stdout is plain text, not a structured event stream; the intermediate "I will run

1. **Make sure the executable is on PATH. ** Open a new terminal and run which <name> (e.g. which claude, which cursor-agent, which kiro-cli, which agy). Print the path and the daemon can find it; if nothing is printed, fix the shell's PATH first (the most common reason is that the rc file is not reloaded).
2. **Restart the daemon process. ** multica daemon restart, or restart the desktop. The daemon scans PATH only once at startup.
3. **Look at the Runtimes page. ** The Runtimes page of the Multica UI should have a line (Workspace × Tools). If "offline" is displayed, see [Daemon and Runtime → When is the runtime marked offline] (/daemon-runtimes#When is the runtime marked offline).
4. **Back to onboarding. ** The "Connect Runtime" step will keep polling, and the new runtime will be scanned within a few seconds without manual refresh.

• **which is found, but not the daemon. ** The daemon was started with the old PATH, restart it.
• **The executable file is there, but the startup fails. ** Run the --version or --help of the tool separately in the terminal. Most of the failures are due to failure to log in, expired token, or incorrect Node.js/runtime version.
• The **Runtimes page shows the line, but the task fails as soon as it is run. ** Run multica daemon logs -f while triggering the task. The daemon process will spit out the tool's own errors as they are.

• Daemon and runtimes — detection, heartbeat, offline processing
• AI Programming Tool Matrix — The difference in capabilities after the tools are connected
• Create and configure agents — Choose a tool for your agent and start running tasks