Coding Horizon

Foxl Makes Every Other AI Agent Feel Backwards

Every figure, name, date and capability this video puts on screen, chased to a primary source. Checked 24 August 2026, against the product’s own site and documentation.

What Foxl is, and where it runs

Foxl ships as three products sharing one account: Foxl Agent (desktop and web), Foxl Code (coding tasks against GitHub repos) and Foxl Notes (meeting transcription).

“Foxl is a 24/7 personal AI agent. It ships as three products that share one account and one credit balance: Foxl Agent …, Foxl Code …, and Foxl Notes …”

Source: https://docs.foxl.ai/llms.txt

Platforms: macOS, Windows and Linux, plus iOS through TestFlight and Android as a direct APK.

Source: https://docs.foxl.ai/docs/get-started/download.md, https://foxl.ai/

The real browser

Foxl drives the user’s actual Chrome, with existing logged in sessions.

“Foxl’s most powerful feature is browser automation using your actual Chrome browser with all your logged-in sessions intact.”

“The Chrome Extension uses your real browser sessions. When you’re logged into Gmail in Chrome, Foxl can read your email.”

Source: https://docs.foxl.ai/docs/desktop/browser.md

Approval gates

“Dangerous actions - shell commands, file writes, browser navigation - require your explicit OK before execution.”

Source: https://foxl.ai/

Pricing and model access

The product is free with every feature unlocked, and the model access is supplied by the user.

“Free, with every feature unlocked. Bring your own model key or run one locally.” “Free forever.”

Three routes are offered: an OAuth subscription (Claude Pro or Max, ChatGPT Plus or Pro, Gemini CLI), the user’s own API key (Anthropic, OpenAI, Google, Amazon Bedrock), or a local model (Ollama, vLLM, LM Studio). Users “pay the provider directly at their published rates”.

Foxl serves no inference of its own: “foxl.ai does not serve model inference. Its hosted catalog is empty and a completion request to it is refused with 409 hosted_inference_disabled.”

Sources: https://foxl.ai/pricing, https://docs.foxl.ai/docs/desktop/providers.md

Foxl Code: the flow

A per user orchestrator turns a conversation into pull requests. It writes a Task Document in markdown; each top level checkbox is one unit of work mapping to one pull request; the user confirms once and the orchestrator then spawns an agent per checkbox.

“A Task Document is the orchestrator’s plan, written in markdown. Each top-level checkbox is one unit of work that maps to one pull request.”

“You confirm once. The orchestrator spawns a coding agent for every checkbox in a single step - they run concurrently, each on its own VM.”

Source: https://docs.foxl.ai/docs/foxl-code/how-it-works.md

Foxl Code: where the agents run, and the six backends

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

The six are Claude Code (default), Kiro, Codex, Cursor, Hermes and OpenCode. Kiro and Cursor run on the user’s own API key.

“Every task runs in a fresh AgentCore micro-VM” with “no state leakage between tasks”, and agents “from different users never share a VM or session.”

Source: https://docs.foxl.ai/docs/foxl-code/coding-agents.md, https://foxl.ai/code

Privacy

Conversations, files, memory and API keys stay on the user’s own hardware; the relay is described as a stateless zero knowledge proxy that forwards encrypted packets it cannot read; local models never leave the machine.

Source: https://foxl.ai/, https://docs.foxl.ai/docs/desktop/providers.md

Memory on disk

Foxl keeps two kinds of memory. Workspace memory is human readable markdown in the user’s workspace directory, and can be read, edited or deleted directly: SOUL.md (injected as the system prompt every turn), USER.md, MEMORY.md, AGENTS.md, TOOLS.md, plus dated daily summaries at memory/YYYY-MM-DD.md. Alongside it there is a key value store held in SQLite, which is not markdown.

Source: https://docs.foxl.ai/docs/desktop/memory.md

Scheduling

Foxl runs scheduled work on the user’s own machine, including a background service that keeps the machine reachable after the app is closed.

Sources: https://docs.foxl.ai/docs/desktop/scheduling.md, https://docs.foxl.ai/docs/desktop/background-service.md

The changelog entry of 24 August 2026

Release v0.6.15, dated 24 August 2026, carries all three items named in the video:

“Hermes is selectable in Foxl Code.” “Drag a workspace file straight into the agent’s prompt.” “Foxl Code: browse your repository’s open issues and pull requests in-app.”

The third adds a Backlog tab showing open GitHub items.

Source: https://foxl.ai/changelog

Caveats

• Memory is not entirely markdown. The marketing site describes agent memory as plain markdown on disk, and the workspace memory files are exactly that. The documentation also describes a separate key value memory held in SQLite. The video says the markdown memory is readable on disk, which is accurate, but “all of Foxl’s memory is markdown” would not be.
• Foxl supplied model access is currently switched off. The Claude Code, Hermes and OpenCode backends are documented as drawing on Foxl supplied model access, which the docs state is switched off at the time of checking. Bringing a key, a subscription or a local model is therefore the route that works today rather than merely one option among several.
• Product maturity. Foxl is at v0.6.x, its iOS build is a TestFlight beta and its Android build is a direct APK. Nothing here is a claim about reliability at scale.