CaoGen 3D office showing multiple AI Agent workstations in progress
v0.1.5 Beta Released
Open source · Provider-independent

CaoGen

Start with a project, then let the right model take each task.

Choose a fixed model, automatic routing within one provider, or automatic routing across providers before a session starts. Execution, review, and delivery remain in the same work desktop.

MIT open source · Current package is unsigned · Verified on Intel Mac

Why CaoGen

More than another chat window

Bring AI into real project work.

CaoGen is built for ongoing projects, not one-off prompts. Project rules, context, execution, cost, and delivery evidence stay on one work chain.

v0.1.5Current beta
3Explicit routing scopes
5Task model roles
33Sub-agent dispatch limit

What's new in v0.1.5

Make the project and model path explicit before work begins.

New sessions no longer ask users to repeatedly choose the underlying engine. Project ownership, routing scope, and task models are selected in one place while compatible native context is preserved during execution.

01

Full-page settings and provider editing

Settings and provider configuration now use a dedicated workspace instead of nested dialogs, making models, keys, protocols, and scheduling rules easier to inspect and maintain.

02

Projects become session containers

Create sessions directly inside a project and keep their context grouped in the sidebar. Unassigned sessions have a separate collection instead of mixing into one history list.

03

Three explicit routing scopes

Choose a fixed provider and model, automatic routing within one provider, or automatic routing across configured providers. Each provider owns its execution engine configuration.

04

Five task-model mappings

Research, planning, development, testing, and documentation can each have a default model and custom rules. Cross-provider selection happens before session creation, with runtime switching limited to compatible engines.

One desktop for the complete workflow

From model selection to delivery.

Each capability participates in real sessions instead of living in a separate showcase panel.

01

Unified multi-provider routing

Manage providers, models, multiple keys, and compatible services in one place, then choose a path by task role, quality, speed, cost, and health.

  • Three session routing scopes
  • Fallback keys within one provider
02

Project sessions and rules

Each project keeps its own sessions, prompts, stack, commands, restricted paths, model policy, and acceptance criteria.

  • Project sessions archived together
  • Rules enter actual model routing
03

Execution and delivery loop

Files, terminal, diffs, Git, worktrees, browser comments, and previews work together inside the task context.

  • Task breakdown and parallel agents
  • Checkpoint recovery and result review
04

Visible, controlled, reviewable

Approvals, routing reasons, budgets, tool state, and task progress remain visible, with evidence retained for important actions.

  • Explicit approval for sensitive actions
  • Continuous runtime visibility

Provider recommendations

Choose third-party API services as needed.

The links below lead to independent provider websites. Each provider determines its actual services, pricing, quotas, and terms.

Third-party API service

CiYuan2API

A registration link for a third-party API service. Available models, pricing, quotas, and terms are determined by the provider.

Visit CiYuan2API Promotional link · CaoGen does not store your provider account or payment information.

Real product interface

Work should not disappear into a black box.

CaoGen desktop with project sessions, model selection, files, plugins, memory, browser, and terminal access
One desktop holds the complete context Sessions, models, files, plugins, memory, browser, and terminal stay in the same workspace.

Core workflow

From project and model selection to delivery.

  1. 01

    Choose the project

    Create a session in an existing project, or keep unassigned sessions in their own collection.

  2. 02

    Choose the routing scope

    Use a fixed model, automatic routing within one provider, or automatic routing across providers before execution begins.

  3. 03

    Classify and execute

    Route by research, planning, development, testing, or documentation preferences, then use files, terminal, browser, and Git.

  4. 04

    Review the result

    Inspect diffs, run verification, and handle actions that require approval.

  5. 05

    Commit and deliver

    Leave reviewable changes, status, and delivery evidence behind.

Multiple Agent workstations and live session states in the CaoGen 3D office

3D office

Not a status animation,
but the actual workplace.

Each session maps to a workstation. Execution, approval waits, recovery, cost, model routing, and worktree state come from real task data.

View the product interface

Get started

Give your next task to CaoGen.

v0.1.5 currently provides macOS x64 and Windows x64 packages. Apple Silicon users can check earlier releases. Linux users can run from source or build locally.

FAQ

Before you download.

Does CaoGen require a Claude login?

No. The default OpenAI-compatible path does not depend on a Claude login. Claude is an optional engine selected explicitly by the user.

Can I connect Chinese or local models?

Yes. Any service with a compatible API can be configured with its provider, model, API key, and custom base URL.

Why does macOS block the first launch?

The current public package is unsigned. Right-click the app and choose Open to confirm launch. Formal signing and notarization are still in progress.

Is it ready for production projects?

CaoGen is currently Beta. The core work path is available for testing. Use a test repository or isolated worktree for high-risk tasks and run your own verification before delivery.

How do the three routing modes differ?

Fixed mode always uses one provider and model. Single-provider automatic mode selects only from that provider. Cross-provider automatic mode selects from configured providers before session creation and only switches to providers compatible with the active engine at runtime, preserving native context.