AI coding tools comparison

Multica does not replace the AI coding tool you use. A runtime invokes the tool already installed and signed in on the computer, then writes progress and results back to the issue.

If you already use one of these tools, you can usually keep it as-is. Four things to weigh when choosing:

• whether it is already installed on the computer;
• which models your account can use;
• whether Multica needs to configure MCP for it;
• which skills the team wants to reuse.

Install steps are in Install AI coding tools.

Supported tools

"Multica-managed MCP" means you can define MCP servers in the agent configuration and Multica passes them to the tool before a run. It does not mean the tool has no other way to configure MCP.

Model sources

When you create an agent, the model list comes from the corresponding runtime. Some tools expose a fixed set of model names; others return available models based on local configuration, the signed-in account, and subscription entitlements.

If you do not pick a model, Multica uses the tool's own default. If the list is empty, confirm the runtime is online and the tool is signed in, then refresh the model list.

QwenPaw is the one exception: it does not accept a model chosen from Multica, so the picker is inert and shows "Managed by runtime". Choose the model in QwenPaw's own configuration instead. An empty model list is expected there and is not a sign that the runtime is offline.

DeepSeek Harness advertises its model catalog from the runtime itself (dsh --profile multica --list-models). Its model ids use the provider/model form (for example deepseek-official/deepseek-chat), so select the full id from the list.

Session resumption

All of these tools support continuing an existing session in a later run. The prerequisites are that the original session still exists and the task is executed by a runtime that can reach it.

Multica stores the session identifier the tool returns. The next time work continues on the same issue, the runtime tries to resume that session; when it cannot, it creates a new session and keeps going. Pi stores a local session file path, so it depends more on the original computer and file.

Retry and session-resumption rules are in Tasks.

MCP configuration

Antigravity, GitHub Copilot CLI, DevEco Code, and Pi do not read the MCP configuration on an agent yet, so the field is not shown when creating agents for them.

Skill injection

Before a run starts, Multica writes the skills bound to the agent into the paths in the table above, so each tool discovers them through its own native mechanism.

Skills that already exist in the repository are never overwritten. On a name collision, Multica uses a different directory name for the injected copy; after the run, it cleans up only the files it created.

Codex, Hermes, and OpenClaw are handled slightly differently:

• Codex uses a per-task CODEX_HOME, so skills are never written into the system-level Codex directory.
• Hermes uses a per-task HERMES_HOME. Existing local configuration is still readable, but bound skills and session data are isolated per task. This isolation kicks in only when the agent has skills bound; tasks without bound skills use the machine's original Hermes home.
• OpenClaw uses skills/ under the current task's working directory, and the per-run configuration sets that directory as the workspace.

For creating and binding skills, see Skills.

Next steps

• Create and configure an agent — model and thinking-effort settings.
• Install AI coding tools — install and detection steps.
• Daemon and runtimes — runtime registration and concurrency.