When Canghe released WeSight on April 8, 2025, the invite codes were being resold for ¥999 within days. The developer community saw a tool that addressed a clear frustration—juggling multiple AI agents across separate terminals and model configurations. But Canghe himself felt the initial version “was not entirely above board,” built on fragile workarounds that couldn’t be sustained. He shut it down after a single day.
Now, after a month of reflection and another month of development, WeSight is back in a fundamentally different form. On June 1st, 2025—coinciding with Children’s Day in China—Canghe open-sourced the entire project under a permissive license. The new WeSight is a desktop AI agent console that acts as a single entry point for managing multiple agent runtimes, from Claude Code and Codex to OpenCode, Hermes Agent, DeepSeek-TUI, and more.
The problem WeSight solves is becoming more acute. According to a 2024 report from RedMonk, developers working with three or more AI coding agents spend roughly 40% of their workflow on configuration and environment switching—tasks that add no direct value. The ecosystem is fragmented: each agent has its own terminal, its own model configuration file, its own plugin system. Tools like cc-switch exist but only handle one agent, and they still require manual terminal intervention. WeSight doesn’t just connect agents—it orchestrates a workflow that was previously buried in terminal commands.
Architecturally, WeSight operates as a virtual harness. It installs and detects local agent environments automatically, and can even bootstrap Claude Code or Codex with a single click. Once attached, all agent tasks are dispatched from the same visual workspace. The model selection layer is unified: a developer can assign the same third-party model (e.g., GPT‑4o, DeepSeek‑V3, Qwen) to multiple agents without editing a single config file. This is a small improvement that removes a major friction point—especially for teams that need to benchmark agents with identical model backends.
The real innovation is the Agent Team abstraction. Canghe observed that previous tutorials on multi-agent solutions like OpenClaw or Hermes Agent were too complex for mainstream adoption—too many steps, too many dependencies. WeSight introduces a higher-level Harness concept where users can choose either a single agent or an Agent Team. This makes multi-agent collaboration (e.g., one agent for code generation, another for testing, a third for deployment) a matter of drag-and-drop selection rather than script orchestration. It is a deliberate design choice aimed at lowering the barrier for developers who are not yet comfortable with YAML pipelines or MCP servers.
A comparison with existing open-source alternatives is instructive. Projects like AutoGPT or AgentGPT focus on autonomous web agents, not local coding agents. LangChain provides infrastructure but no desktop GUI. Claude’s own desktop app is limited to a single agent. WeSight fills a specific niche: a local-first, visual control panel for the growing list of coding agents that developers already run. And because it’s open-source, the community can extend it—adding new runtimes, custom tools, or integration with local model providers like Ollama and vLLM.
Canghe’s decision to open-source after the initial controversy is itself noteworthy. The earlier version operated in a legal gray area, likely by scraping or redistributing agent binaries without authorization. The new WeSight is built on official APIs and licensed runtimes, making it “legitimate” enough for public release. This pivot reflects a broader trend in the AI tools ecosystem: developers are willing to pay for convenience, but they also demand transparency and compliance. Trust is a dependency that no agent can auto-install.
Looking ahead, the success of WeSight will depend on its ability to keep pace with the rapid release cycle of coding agents. Claude Code updates weekly; Codex merges daily. Maintaining a harness that doesn’t break on every update is a nontrivial engineering challenge. But the open-source model mitigates this: early adopters can submit patches, and the community acts as a distributed QA team.
For now, WeSight offers a glimpse of a future where developers no longer need to remember six different terminal commands to switch between agents. It provides a single window, a single model configuration, and a single task queue. Whether that future arrives depends on how many developers are willing to step out of their terminal comfort zone. But given the ¥999 hype that preceded the shutdown, the demand for simplicity is clearly there.