CZL LogoCZL
  • Home
  • Worldwide Express
  • 中文
Built in-house · Agent Harness

CZL AI

Act. Create. Build.

More than a chat box. CZL Agent plans, splits up work, calls tools and checks its own work on an in-house harness until the job is done; AI apps work out of the box; the API gateway gives you every major model through one endpoint.

Pick your systemDownload desktop app
No install neededOpen in browser
Download installerAndroid
Desktop app for Windows and macOS plus an Android app; the same account works on the web and as a PWA
CZL AI
CZL
Bottle launch
Competitor research · YETI
Competitor research · Hydro Flask
Q3 overseas ads review
Supplier quote comparison
Our insulated bottle launches next month: write the launch plan, research three competitors (YETI, Hydro Flask, Stanley), and build a budget Excel and a 5-slide PPT.
Web search · 6 times
Run code · files generated

Stanley has little public material, so I'll look into it myself; the other two go to sub-conversations running in parallel, and they'll report back when done.

Task list2/4
  • Draft the launch plan
  • Build the budget Excel
  • Research 3 competitors
  • Check PPT figures against findings
Competitor research · YETIRunning
Competitor research · Hydro FlaskReported
Bottle_Launch_Budget.xlsx
Launch_Plan_5_Slides.pptx
Type to continue — you can interject mid-run…
Environment
This computer · Bottle Store
Context
18%
Changes
+42 −8 · 3 files
Background tasks
2 sub-conversations
Outputs
Bottle_Launch_Budget.xlsx
Launch_Plan_5_Slides.pptx
CZL AI desktop interface sketch: sub-conversations in the sidebar, a task list, output files and the right-hand side panel

01Platforms

Three platforms, one account

One balance and one plan across web, desktop and API. Create an organization to share a balance with your team, with monthly limits per member.

CZL Agent

An agent that acts: searches the web and reads pages directly, writes and runs code, generates files, remembers you, sets reminders — and on desktop it works on your computer.

AI Apps

Ready-made tools — Imagic (with GIF animation, SVG vectors and draw-on-image markup edits), background swap, 5-shot e-commerce images, multi-style images, email replies. Fill in a form, get the result.

API Gateway

An OpenAI-compatible unified endpoint with Responses / Chat Completions / Messages protocol conversion and automatic multi-upstream failover. Supports OpenAI Fast / Ultrafast tiers and Claude fast mode, billed by the tier actually served.

02Multi-agent

One person, a small team

On desktop, CZL Agent hands independent subtasks to sub-conversations that work in parallel on your own computer — reading and writing project files and running commands. Open any of them to watch, or add requirements at any time.

  1. 01

    Propose first, then act

    When it wants to split work off, the agent shows a proposal card describing the job. You click Start, or choose Do it here — the call is yours. When useful, a sub-conversation starts with the main conversation's context, so nothing needs retelling.

  2. 02

    Work in parallel

    Each sub-conversation is a real conversation under the main one, using the same local tools to finish its own task. For code changes you can pick a separate copy: it works on its own Git branch, so several sub-tasks can edit at once without overwriting each other.

  3. 03

    Automatic report-back

    When a sub-conversation finishes, it reports to the main conversation and wakes it up; the main one checks the result and carries on — no need to babysit.

  4. 04

    Step in any time

    Open a sub-conversation to see the whole process or add a requirement; the main conversation can send it follow-up instructions too.

Desktop only. The web agent also has task lists and clarifying questions, but cannot spawn sub-conversations.

03Meet Mu

Hand it a task. It gets it done.

Mu is CZL's border collie mascot. A collie gets the whole flock where it needs to go on its own, and CZL Agent works the same way: it plans, acts and delegates until the job is finished.

About Mu
Try CZL Agent

04Capabilities

What a plain chat page can't do

The agent doesn't just generate text — it delivers working results.

01

Code generation and editing

Bind a local project folder and it reads, writes and edits code, runs builds and tests. Each edit shows up in the answer with a line-by-line diff; scripts run instantly in the code sandbox; on desktop, click a local file path in an answer to open it.

02

Word / Excel / PowerPoint and files

Generates documents, spreadsheets, slides and charts with Python in the sandbox, delivered as attachments you can keep refining in later turns.

03

Talk with images

Understands uploaded images, generates and edits images in the conversation including masked inpainting, short looping GIFs and editable SVG vector graphics such as logos and icons — with Claude, Gemini or any other model, and as one step of a larger task; on desktop it can open images on your computer and save generated images and files to any folder you choose, ready for local scripts and tools; past images are replayed across turns.

04

Seamless with Claude Code / Codex

Discovers the MCP servers and Skills you already set up in Claude Code and Codex, and applies your project's CLAUDE.md / AGENTS.md as rules — your dev environment works out of the box.

05

Task list

For complex work the agent writes its own plan and ticks items off as it goes, and periodically looks back at it mid-run so it doesn't drift. The list is saved with the conversation and survives a refresh.

06

Clarifying questions

When something is unclear it asks you instead of guessing. Multiple questions are answered page by page, and any of them can be skipped.

07

Right-hand side panel

Environment and context usage at a glance, with outputs (files and images the AI generated) and background tasks in one place. On desktop you can also browse project files, open a terminal in the project folder, review Git changes and switch branches without leaving the chat.

08

Queueing, steering and a real stop

Messages sent mid-generation show up in the conversation right away, images included, and are injected between tool calls so you can course-correct; stop really stops, and an interrupted answer resumes from where it left off. Your input draft and pending messages survive closing the window or refreshing.

09

Work in your own browser

Install the "CZL Agent" extension for Chrome / Edge and the AI can work in the browser you already use, signed in as you: open pages, read them, fill in forms. Tabs it opens go into a "CZL Agent" tab group, and it asks first before anything irreversible: submitting, paying, deleting, entering passwords, running scripts or touching your own tabs.

10

Use desktop software like a person

Hand it apps that have no scripting interface: it takes screenshots, clicks, types, switches windows and reads on-screen controls. Off by default, asks once per conversation before the first mouse or keyboard action, waits while you are using the mouse, and can be stopped any time. Windows and macOS. Office, Photoshop, AutoCAD, Blender and SketchUp also get built-in skills you can enable on demand.

05Architecture

In-house Agent harness

The loop, memory, context, tools and permissions around the model, built in-house on frontier agent-engineering theory.

Harness framework

01
  • A perceive–plan–act–observe agentic loop where the model decides the next step
  • No cut-off by round count or time; it only moves to a wrap-up round when the agent is going in circles (the same tool with the same arguments, again and again) or the context is nearly full and cannot be compacted, so there is always a deliverable
  • Every round is checked against your balance again, and the run pauses and tells you when it runs short
  • Human-in-the-loop: high-risk actions pause for your confirmation
  • Reasoning continuity: thinking stays linked to the tool calls it led to, across rounds

Four-layer RAG memory + project experience

02
  • Working memory: live state and situational awareness of the current session
  • Episodic memory: summaries of past sessions, recalled by semantic relevance
  • Profile memory: a stable long-term user profile that stays in context
  • Semantic memory: a vector fact store for retrieval-augmented generation, with automatic extraction, dedup and forgetting
  • Project experience: it records conventions, commands and pitfalls as it works in a project, with an index always in context and notes read on demand; it adds useful files and images to the project knowledge base itself, so each project gets smoother the more you use it
  • Project organization: with your approval, the agent can create a project and move the current conversation into it, move it to an existing project, or move it out. The new project environment applies next turn; local folders are still bound through your folder picker

Context engineering

03
  • Uses the full context window of the model you pick (up to about 1M tokens), enough room for long tasks
  • High-signal, low-noise context supply
  • Prompt-cache-friendly layered assembly that keeps stable and volatile content apart
  • Progressive context compaction: it only compacts when the context is nearly full, including mid tool loop, then carries on, so long tasks never lose the thread; you can also compact manually at any time
  • Progressive disclosure of tools, so a large toolset never crowds the context

Planning and multi-agent

04
  • Task list and plan checks: the list is stored with the conversation, and reminders are injected periodically mid-run to keep work on plan
  • Sub-conversation orchestration: a single level only, sub-conversations cannot spawn more; they run on the same computer as the main conversation with read-only tools by default and write access only when needed; can start from the main conversation's context (fork); a cap on how many run at once
  • Isolated parallel writes: a sub-conversation can work in its own Git worktree branch for the main conversation to merge; in a shared project folder, a write lease keeps parallel tasks from overwriting the same file
  • Asynchronous report-back: a finished sub-conversation reports to the main one and wakes it; if sibling sub-conversations are still running, results are delivered together once the last one ends
  • Unattended continuation: a limit on consecutive automatic turns, auto-wake pauses after you stop a run, and unattended turns can only propose work, never start it on their own

Code sandbox

05
  • Isolated execution of untrusted code
  • Output as delivery: execution results become downloadable files
  • Closed-loop self-correction driven by execution feedback

Local tools with a bundled runtime

06
  • Least privilege with tiered authorization
  • Execution decoupled from the UI: tasks live on business entities and survive reloads and closed windows
  • Batteries included: a bundled runtime that works with zero setup
  • Ecosystem-native: compatible with the MCP protocol and the Skills spec

Reliability engineering

07
  • Cooperative cancellation: stop means really stop
  • Instruction hierarchy: web pages, files and tool output are data, not instructions
  • Idempotency and honest state: verify when a result is unknown, never repeat side effects
  • Streaming fault tolerance: pre-first-token buffering, multi-route failover and graceful shutdown

06In-house

Built in-house, on frontier ideas

CZL Agent's harness, four-layer memory, context engineering, tool scheduling, chat frontend and Wails desktop app are all developed in-house by CZL; the API gateway's protocol conversion, billing and failover were redesigned by CZL as well. The design draws on the published ideas and engineering experience of OpenClaw, OpenAI Codex and Anthropic Claude Code, and keeps iterating on real production data.

  • No third-party agent framework — the loop, tools and memory are built layer by layer
  • Draws on Codex's state-consistency design and Claude Code's permission levels and Skills
  • Stop, retry and idempotency designed after OpenAI / Anthropic agent engineering practices

Get started

The desktop app unlocks local tools, project folders (files, terminal and Git), Skills and MCP; the Android app keeps replying in the background and notifies you when it is done; the web app is ready instantly.

Pick your systemDownload desktop app
No install neededOpen in browser
Download installerAndroid
API docs

CZL AI is developed and operated by CZL·About CZL

CZL focuses on two business lines: AI and international logistics.

CZL LogoCZL

Cross-industry AI and dependable international logistics for global business

CN · SG · UK · US / EST 2020

AI Products

  • CZL AI
  • CZL TransAI
  • Imagic

Productivity & Dev Tools

  • czlterm
  • CZL Mail
  • CZL Browser

Trade & Logistics

  • CZL TradeDocs
  • International Express API

Online Services

  • CZL Image Hosting
  • CZL Connect
  • Navigation

About

  • About CZL
  • Mu, Our Mascot
  • Contact Us
  • Service Status
  • Terms of Service
  • Privacy Policy
Copyright © 2026 CZL All Rights Reserved.Company Information
沪ICP备20015193号沪公网安备 31011702008265号