Coding agents

A coding agent is a real coding CLI with a filesystem and terminal. Remote tasks run on isolated Amazon Bedrock AgentCore machines. Claude Code, Codex, Hermes, and OpenCode use managed Relay credits; Kiro and Cursor use your vendor API key and charge Foxl credits for compute only.

Cloud backends앵커 링크 복사

Claude choices include Opus 5.5, Opus 5, Fable 5.1, Fable 5, Opus 4.8, Sonnet 5.5, Sonnet 5, and Haiku 5.5. Codex choices include GPT-6 Astra, GPT-5.6 Sol / Terra / Luna, and GPT-5.5. GPT-6 Sol and Luna are available in the desktop app with supported personal providers, but are not on managed Code credits. Older versions remain available through the model preferences. A chat-model choice does not change which models a coding CLI supports.

Remote tasks run Claude Code 2.1.296, Codex 0.162.1, Kiro CLI 2.29.1, Cursor Agent 2026.10.01, OpenCode 1.19.0 and Hermes 0.19.0.

• Claude Code on Opus 5.5 gets the model's full 1M-token context on Remote, so a long task is not summarized at 200K. Every other Claude model works in a 200K-token context there, and a task on a machine of your own runs whatever CLI is installed on it. A turn with a very large context needs more credits set aside before it runs; see Budgets and tiers.
• Codex defaults to GPT-5.6 Sol, and runs GPT-6 Astra with that model's own settings when you pick it. When your account's simultaneous model requests are all in use, it waits and retries a few times before it fails. It can search the web with Amazon Bedrock Web Search, which is charged per search.
• The cloud machines include the sqlite3 command, so a repository's test suite that calls it can run there.

Choosing an agent앵커 링크 복사

• Per task: choose a backend when creating a task, or name a supported backend in chat.
• Account default: use the coding-agent picker in the Tasks message box, the agent segment of the Runtimes header, or Settings > Code > Coding agent. This also applies to new scheduled and autonomous tasks.
• Terminal: opens a shell without a coding assistant. Remote compute is metered even when no model is running.
• Models and credits: managed agents use the model picker without a personal key. Save a vendor API key in Code settings before starting Remote Kiro or Cursor.

An existing task keeps its agent and model when resumed. Only Claude Code tasks can be resumed or reconnected once their agent has exited; Codex, Kiro, Cursor, Hermes and OpenCode tasks use Retry.

Desktop local agents앵커 링크 복사

The desktop app's local execution path keeps its own CLI credentials and provider billing. Supported local choices include Claude Code, Codex, Kiro, Cursor, Hermes, and OpenCode. See On your own computer.

Personal credentials saved by older versions of Code are not used for managed cloud tasks. Changing a provider in the desktop app does not change the cloud task's model access.

Isolation and the workspace앵커 링크 복사

Each Remote agent gets its own session. Its checkout and build artifacts live on the machine, which is recycled after fifteen idle minutes and has an eight-hour maximum lifetime; a reconnect clones the pushed branch back, and in a Claude Code terminal pane on a private repository puts back the uncommitted changes saved to a ref under refs/foxl/wip/ in your repository every two minutes (never a file that holds one of the task's own credentials). A task that runs in the background gets no such save yet (see Resume from review). Claude Code's conversation is kept in session storage for 14 idle days. Keep durable work in a pushed branch.

Configuring the container앵커 링크 복사

Settings > Code > Container controls:

MCP servers you add under claudecode in the Advanced config are written to the file Claude Code reads them from, so a cloud Claude Code task with Work without asking permission on starts them; your other Advanced settings for Claude Code stay in its settings file. There is no time-limit setting: see Tasks for how long a run can go. These settings are reapplied on reconnect. Repository instruction files are preserved. Managed model IDs, provider endpoints, and authentication are controlled by the task and cannot be replaced by personal credentials in container settings.

Managed model usage goes through the task's Relay authorization. Kiro and Cursor use their vendor billing; separate orchestration or reviews can still use managed models. See Budgets and tiers for how credits are reserved and settled.