Coding agents

Claude Code, Kiro, Codex, Cursor, Hermes, and OpenCode backends and how to choose one

Coding agents

Each task is carried out by a coding agent - a real coding CLI running in an isolated AWS Bedrock AgentCore microVM with a full filesystem and terminal. Foxl Code ships six interchangeable backends.

Backends

BackendCLIModelsKey
Claude Code (default)Anthropic Claude CodeOpus 5 (default), Fable 5, Opus 4.8 / 4.7 / 4.6, Sonnet 4.6, Haiku 4.5Included
KiroKiro CLIClaude familyBring your own
CodexOpenAI Codex CLI on Amazon BedrockGPT-5.6 Sol / Terra / Luna, GPT-5.5, GPT-5.4Included
CursorCursor CLICursor picks the model (auto)Bring your own
HermesHermes CLIClaude family on BedrockIncluded
OpenCodeOpenCode CLIClaude family on BedrockIncluded

All six run the same way - inside an interactive shell in their own VM - so the live terminal, resume, and budget behavior are identical regardless of which one you pick.

Choosing a backend

  • Per task - name a backend when you create the task and the orchestrator honors it.
  • Default - set your preferred coding-agent family and, for Codex, the model (GPT-5.6 Sol / Terra / Luna, or GPT-5.5 / GPT-5.4) in Settings. Tasks with no explicit backend use this default.
  • Bring-your-own-key backends - Kiro and Cursor use your own API key, added in Settings. Their model usage is billed to your key, not your Foxl credits; the AgentCore compute time is still metered normally.

Backends not provisioned in your environment are greyed out in Settings. The available model list also depends on your plan - see Budgets and tiers.

Isolation and the workspace

Every agent runs in its own microVM with a working directory at /mnt/workspace. The filesystem is session storage: it survives across a task's invocations and idle gaps, so retrying a task or resuming from review keeps the cloned repo, installed packages, and any build artifacts intact. Agents from different users never share a VM or session.

Models

Coding-agent model availability tracks your plan:

  • Claude Code / Hermes / OpenCode - the Claude lineup (Opus 5, Fable 5, Opus 4.8, Sonnet, Haiku); Foxl-billed Claude models require Pro or higher. Opus 5 is the coding-agent default: it matches or beats Fable 5 on agentic coding and professional knowledge work at half the price, with adaptive thinking on by default, a 1M-token context window, and up to 128K output tokens. Opus 4.8 stays selectable. Opus 5 needs Pro, Ultra, or Enterprise; the free tier keeps Sonnet and Haiku.
  • Codex - GPT-5.6 Sol (most capable), Terra (best price-performance), and Luna (fastest), plus GPT-5.5 / GPT-5.4, served through Amazon Bedrock with no API key of your own.
  • Kiro / Cursor - run on your own API key, so Foxl's model tier rules do not apply. Cursor selects the model automatically; there is no per-task model picker.

The model your coding agent uses is separate from the model the orchestrator uses for chat replies.

On this page