AppX Labs

Independent software studio · Taiwan

Agents you can trust with real work.

AppX Labs builds AppX Harness — an agent harness that wraps Claude models with tools, MCP, skills, memory, human approvals, and observability, so a solo builder or small team can put agents into real work.

How it works See the roadmap
run trace · example
  1. planTriage new support emails, draft replies
  2. toolinbox.search (MCP)ok
  3. skillsupport-triageloaded
  4. memoryproject › tone & policiesread
  5. tooldrafts.create ×3ok
  6. approveemail.send ×3waiting for you
Illustrative trace. Reads run freely; outbound actions wait for a human.

AppX Harness

Everything around the model that makes an agent dependable.

Claude does the reasoning. The harness supplies the rest: what the agent can touch, what it remembers, when it must ask, and a record of everything it did.

Tools & MCP

Connect Claude to APIs, files, and apps through Model Context Protocol servers and typed tools, each with its own scopes and limits.

Skills

Package repeatable know-how — instructions, scripts, and references — as skills the agent loads only when a task calls for them.

Memory

Durable, inspectable memory across sessions: project notes, preferences, and task state that you can read, edit, or reset.

Approvals

Human-in-the-loop by policy. Low-risk reads run on their own; sends, writes, payments, and deletes pause until a person approves.

Observability

Every run is traced — prompts, tool calls, approvals, tokens, cost, and latency — so you can replay failures and turn them into evals.

Reliable runtime

Background jobs, schedules, retries, timeouts, and sandboxed execution, so long-running agents finish their work or fail loudly.

Built with Claude

Claude in the product, and in how we build it.

We chose Claude for its tool use, long context, and dependable instruction following — the qualities that matter most when a model is trusted to act.

In AppX Harness

  • Claude API as the model layer. The harness is built on the Messages API with tool use, streaming, and prompt caching.
  • Right model per step. Larger Claude models plan and handle hard steps; faster ones take routine tool calls and summaries.
  • MCP-native. Tools are exposed through the Model Context Protocol, so existing MCP servers plug in directly.
  • Skills and memory designed for Claude. Context is loaded progressively to keep prompts focused and costs predictable.

In our development process

  • Built with Claude Code. AppX Harness is developed with Claude Code: writing, refactoring, reviewing, and testing the codebase.

Status

Early stage, building in public.

AppX Harness is in active development. Early access for a small group of builders will open in a later phase.

  1. Now

    Core runtime

    Claude tool-use loop, MCP connections, approval policies, and run tracing.

  2. Next

    Skills & memory

    Skill packaging, persistent memory with an editor, and a trace viewer for debugging runs.

  3. Later

    Early access for small teams

    Hosted workspaces, shared approvals, and evals built from production traces.

About

A one-person studio for the agent era.

AppX Labs is an independent, one-person software studio based in Taiwan. We build developer tools for working with AI agents, starting with AppX Harness.

We believe agents become useful when they are dependable: scoped access, clear memory, a human in control of consequential actions, and a full record of what happened. That is what we are building.

  • Humans stay in control

    Agents propose; people approve what matters.

  • Observable by default

    If an agent did it, you can see it and replay it.

  • Small and dependable

    Fewer moving parts, honest defaults, shipped often.