Foxl Code
Autonomous coding agents at code.foxl.ai
Foxl Code
Foxl Code is a fleet of autonomous coding agents running on AWS Bedrock AgentCore. You describe work in natural language, and the orchestrator plans, spawns agents, and manages their lifecycle - from opening GitHub Issues to pushing branches and creating PRs.
How it works
- Chat with the orchestrator at code.foxl.ai. Describe what you want done.
- The orchestrator creates a Task Document - a markdown plan with checkbox subtasks.
- For each subtask, it spawns a coding agent on an isolated cloud VM.
- Each agent runs Claude Code CLI with full terminal access. You see real-time PTY output (colors, spinners, syntax highlighting) in your browser.
- When agents finish, they push branches and open PRs. If they need your input, they pause in "review" state.
Key concepts
Tasks
A task is one unit of work. It has a prompt, a budget, a model, and a lifecycle:
- queued - waiting for a container
- running - agent is active
- review - agent paused, waiting for your input
- merged - PR merged, work complete
- failed - agent errored out
- cancelled - you stopped it
Task Documents
A Task Document is the orchestrator's plan. It contains subtasks as checkboxes. As agents complete work, checkboxes get ticked. You can see the plan in the right sidebar on the Tasks page.
Autonomous Loop
The orchestrator can check back on running tasks automatically. Enable it in Settings:
- Interval - how often to check (default 5 minutes, minimum 1 minute)
- Quiet hours - time range when the loop sleeps (e.g. 22:00-06:00)
- Daily cap - maximum USD spend per day on autonomous ticks (default $0.50)
The loop only fires when something has actually progressed (new events since last check). Idle users cost nothing.
Budgets
Every task has a spending limit. The relay tracks LLM costs in real time:
- Auto-estimate - based on model, expected duration, and vCPU/memory profile
- Custom budget - set before launch in the composer's "Budget (USD)" field
- Multiplier - per-user scaling factor in Settings (e.g. 2x for complex repos)
When a task hits its ceiling, the agent commits what it has and stops. No overcharges.
Tiers
| Tier | Credits/mo | Models | Concurrent agents |
|---|---|---|---|
| Free | 10 | Haiku, Sonnet, GLM-5, Kimi | 5 |
| Pro ($20/mo) | 500 | All (incl. Opus) | Unlimited |
| Ultra ($200/mo) | 10,000 | All + priority | Unlimited |
| Enterprise | Usage-based | All + dedicated infra | Unlimited |
Credits fund both LLM tokens and AgentCore compute. Conversion: 1 credit = $0.04 USD.
Connecting GitHub
- Go to Repos in the sidebar
- Click Connect GitHub
- Install the Foxl Bot GitHub App on your account or org
- Select which repositories to grant access to
- Your repos appear in the Repos page and the topbar repo picker
The GitHub App gets: Contents (R/W), Issues (R/W), Pull Requests (R/W), Metadata (R). Tokens are minted per-task and never stored.
Resume from review
When an agent pauses for review:
- Read its output on the task detail page (full terminal history)
- Optionally send input via the text field (typed directly into the agent's PTY)
- Click Resume to continue from exactly where it stopped
If the container was reaped (idle timeout), the UI automatically falls back to a retry with the same session storage (preserving git state and installed packages).
Agent interaction
The orchestrator can:
- Read an agent's screen - see what it's currently showing
- Send input - type into its terminal
- Resize - adjust terminal dimensions
- Poll active agents - check which agents are still running
This enables multi-agent coordination where the orchestrator monitors several agents and unblocks them as needed.