WeSight

WeSight

by freestylefly (苍何)

Open-source desktop control console that puts a graphical workspace in front of terminal coding agents — Claude Code, Codex, OpenClaw, Hermes Agent and others — with unified model routing, MCP management and runtime telemetry.

See https://wesight.ai/ · source at https://github.com/freestylefly/wesight

What it actually is

WeSight is not another coding agent. It is a shell around the agents you already run. The premise is that terminal-native agents are individually capable but each keeps its own setup, model config, permissions, IM hooks and logs in a different place; WeSight consolidates those into one Electron desktop app and gives them a shared visual surface.

It either installs a supported CLI for you or detects and reuses an existing local install, so it sits on top of your current setup rather than replacing it.

Supported engines

Engine	Integration path
Built-in runtime	Bundled — uses @anthropic-ai/claude-agent-sdk
Claude Code	One-click install or reuse local CLI
Codex	One-click install or reuse local CLI
Kimi Code	Official installer or local CLI; uses @moonshot-ai/kimi-agent-sdk
OpenClaw	Local runtime reuse or WeSight setup flow
Hermes Agent	Official installer or local CLI
OpenCode	One-click install or reuse local CLI
Qwen Code	One-click install or reuse local CLI
DeepSeek-TUI	One-click install or reuse local CLI

Features

• Unified model providers — OpenAI, Anthropic, Google Gemini, DeepSeek, Qwen, Moonshot, Ollama, OpenRouter, GitHub Copilot, and any OpenAI-compatible endpoint, configured once and mapped into whichever engine is selected
• Graphical tool execution — commands, file diffs, permission prompts, slash commands, generated images and tool results rendered inline in the chat rather than scrolling past in a terminal
• MCP management with a GUI — MCP servers stored in SQLite with stdio, sse and http transports, added and toggled through a form rather than hand-edited config
• AI Runtime Dashboard — per-call telemetry: engine, model, tokens (including cache read/write), TTFT, output-phase TPS, estimated model TPS, tool latency, agent steps, cost and status
• IM Agent Hub — routes Feishu messages into OpenClaw, Hermes, Claude Code or Codex with per-engine bot profiles; ships @wecom/wecom-aibot-sdk for WeCom
• SkillHub marketplace — install, enable, update and remove skills; 20 ship in-repo (article-writer, pdf, pptx, playwright, remotion, seedance, seedream, skill-creator, skill-vetter, imap-smtp-email, …)
• Scheduled tasks — recurring agent jobs for research, reports, monitoring and inbox cleanup
• Memory and personalization — extracts preferences from conversations and reuses them in later sessions
• Live Workspace — right-hand panel showing live file writes, static diffs, todos, skills and artifacts as the agent works
• Desktop pet and pixel studio — a lightweight companion that tracks active tasks

Architecture

Electron app, TypeScript throughout, ~443 source files and roughly 152k lines. Local state in SQLite via better-sqlite3.

The interesting part is src/main/libs/agentEngine/ (~12k lines), a runtime-adapter layer. Every engine implements the same event contract — message, messageUpdate, permissionRequest, runtimeMetric, complete, error, sessionStopped — and a router dispatches to the right adapter.

flowchart TD  
    UI["Electron renderer: chat, tools, dashboard, SkillHub"] --> ROUTER[coworkEngineRouter]  
    ROUTER --> A1["claudeRuntimeAdapter - 83 lines"]  
    ROUTER --> A2["codexAppRuntimeAdapter - 1.1k"]  
    ROUTER --> A3["hermesRuntimeAdapter - 939"]  
    ROUTER --> A4["openclawRuntimeAdapter - 4.3k"]  
    ROUTER --> A5["deepSeekTuiRuntimeAdapter - 647"]  
    ROUTER --> A6["externalCliRuntimeAdapter - 3.6k, generic CLI path"]  
    A1 --> PROV["Unified provider config"]  
    A6 --> PROV  
    PROV --> MCP["mcpStore - SQLite, stdio/sse/http"]  
    ROUTER --> TEL["Runtime telemetry: tokens, TTFT, TPS, steps"]  

Two observations from reading the source that the README does not state:

1. The abstraction is Claude-shaped. agentEngine/types.ts imports PermissionResult directly from @anthropic-ai/claude-agent-sdk, so the common interface every other engine is adapted into is Anthropic’s permission and event model. claudeRuntimeAdapter.ts is only 83 lines because Claude Code needs almost no translation; openclawRuntimeAdapter.ts is 4,262.
2. Adapter size is a fair proxy for integration depth. OpenClaw and the generic external-CLI path account for two-thirds of the adapter layer. Engines routed through externalCliRuntimeAdapter get a shallower integration than the ones with a dedicated adapter.

Superpowers

The real draw is consolidation for people who run several agents. If you use one agent, its own CLI is fine. If you run Claude Code and Codex and OpenClaw against different models, WeSight collapses provider config, MCP server management, permissions and cost tracking into one place — and the runtime dashboard is more detailed than what any of those CLIs report natively.

The second draw is making agent work legible to non-terminal users. Rendering permission prompts, diffs and tool output as UI rather than scrollback lowers the barrier for people who would not adopt a TUI, which is also what the Feishu/WeCom routing is for: driving agents from chat apps.

Ecosystem

Repo	Stars	License	Notes
freestylefly/wesight	904	MIT	The desktop app. v1.0.6, active
freestylefly/wesight-obsidian	365	AGPL-3.0	Runs Claude Code, Codex, OpenCode inside an Obsidian vault
freestylefly/wesight-cli	7	MIT	Agent Runtime Gateway CLI; untouched since May 2026

The Obsidian plugin is the directly relevant one for this vault — sidebar chat, inline editing of selected text, file mentions, model-provider profiles, slash commands, and local conversation storage in .wesight/. Note it is licensed differently from the desktop app and reaches api.wesight.ai for sharing, WeChat Official Account drafts and “Creator member” features, so parts of it are not self-contained.

Caveats

• Windows installer is unsigned — SmartScreen will warn; verify against SHASUMS256.txt. macOS builds are signed and notarized for both Apple Silicon and Intel.
• Chinese-first project. CHANGELOG.md and IDENTITY.md are Chinese-only; README_zh.md is the fuller document. Feishu and WeCom are first-class where Slack and Discord are absent.
• Young and moving fast — created 2026-05-07, at v1.0.6 by 2026-08-24. The changelog’s own “known issues” section admits build/lint/test results were not re-recorded for a release branch.
• Identity has churned. package.json still carries "description": "AI-Powered WeChat Intelligence", and the changelog records rebranding built-in skills “from LobsterAI to WeSight”. The WeCom SDK dependency and IM-first features are residue from that earlier product; it explains the shape but is worth knowing before judging the roadmap as settled.
• “Open source” is not uniform across the ecosystem — MIT desktop app, AGPL plugin, cloud dependency for some plugin features.

Pricing

Desktop app is free and MIT-licensed; you pay only for the model API keys you supply. Signed macOS and Windows builds are published on GitHub Releases at no cost. The Obsidian plugin gates some features (Knowledge Brain, sharing, WeChat publishing) behind a WeSight account and “Creator member” status.

Related

• Claude Code vs Codex vs Goose vs Hermes vs Pi — WeSight fronts three of those five; this note is the console, that one is the agents
• Claude Code · Codex · OpenClaw · Hermes Agent · OpenCode — supported engines
• Claude Agent SDK — the built-in runtime and the shared event contract are built on it
• Model Context Protocol — managed through a GUI here rather than config files
• Obsidian — companion plugin runs agents inside the vault