ChatClaw is a Go desktop agent with an OpenClaw style skill market, and its server mode defaults to all interfaces
ChatClaw: Get OpenClaw-like knowledge base personal AI agent in 5 mins. Sandbox-secured, ultra-small 30MB installer for macOS & Windows (install in 1 min). Connects to WhatsApp, Telegram, Slack, Discord, Gmail, DingTalk, WeChat Work, QQ, Feishu. Built-in Skill Market, Knowledge Base, Memory, MCP, Scheduled Tasks. Developed in Go ,run
At a glance
- What is it?
- ChatClaw is a Wails v3 desktop application, written in Go with a Vue 3 frontend, that puts a local knowledge base, a 5000 plus OpenClaw skill library, multiple configured agents, and scheduled tasks behind a chat interface, with bridges to WeChat, DingTalk, Feishu, QQ, and WhatsApp. The part worth reading carefully is deployment: server mode binds 0.0.0.0:8080 by default and the release links point at a different GitHub organisation than the repository itself.
- Who is it for?
- ChatClaw fits a Windows or macOS user who wants an OpenClaw style agent with a document knowledge base and no scripting, especially if the work arrives over WeChat, DingTalk, Feishu, QQ, or WhatsApp. It does not fit a headless Linux deployment without reading the networking defaults, because server mode listens on all interfaces unless you override it.
- Can I use it commercially?
- Yes, with conditions. GPL-3.0 is a copyleft licence: if you distribute software that includes it, you must release that software's source code under the same licence. Running it internally without distributing it does not trigger that obligation.
- Is it still maintained?
- Activity is slowing. The repository last received commits 6 months ago.
- What is it written in?
- Mainly Go, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- zhimaAi/ChatClaw316 ★ · +8
| Repository | 2026-09-13 | 2026-11-15 |
|---|---|---|
| zhimaAi/ChatClaw | 308 | 316 |
8 Stars in the last 63 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Server mode binds 0.0.0.0:8080 unless you override it
ChatClaw can run without the desktop GUI and be driven from a browser, which is how you would put it on a small server. Two Linux binaries are published, `ChatClaw-server-linux-amd64` and `ChatClaw-server-linux-arm64`, and the run sequence is short:
chmod +x ChatClaw-server-linux-amd64
./ChatClaw-server-linux-amd64Then you open `http://localhost:8080`. The part that deserves attention is what the service binds by default: it listens on `0.0.0.0:8080`, not on loopback, so anything that can reach the host can reach the interface. On a workstation that is harmless, and on a shared host or a cloud instance it is an open web application unless a firewall or proxy stops the traffic first.
The override is an environment variable pair, which is the documented fix:
WAILS_SERVER_HOST=127.0.0.1 WAILS_SERVER_PORT=3000 ./ChatClaw-server-linux-amd64Note that the variable names still carry the Wails prefix even in a product whose value proposition is being OpenClaw-like. That prefix is the desktop framework it is built on, and it leaks into the deployment surface.
Persistence follows the same convention, with a named volume mounted at `/root/.config/chatclaw` so the database and configuration survive a container replacement.
The release links point at a different GitHub organisation than the repository
The server mode instructions tell you to download binaries from `github.com/chatwiki/chatclaw/releases`, and the Docker image is pulled from `registry.cn-hangzhou.aliyuncs.com/chatwiki/chatclaw:latest`. The repository you are reading, however, is under the `zhimaAi` organisation, and the project homepage is `ichatclaw.com`.
Three names, one product. That is not necessarily wrong, since a project can publish artefacts from an organisation that holds release infrastructure, and the Chinese name for the project is different again. But it means the thing you install and the thing you read are not at the same address, which is worth confirming before you run an installer on a work machine.
The Docker image is also hosted on an Alibaba Cloud registry rather than Docker Hub, which is a deliberate choice for users in that region and a source of friction elsewhere.
The container run line is otherwise conventional:
docker run -d \
--name chatclaw-server \
-p 8080:8080 \
-v chatclaw-data:/root/.config/chatclaw \
registry.cn-hangzhou.aliyuncs.com/chatwiki/chatclaw:latestA compose file is provided for the same layout, with `docker compose up -d` to start and `docker compose down` to stop, and data kept in the `chatclaw-data` volume.
The repository's compose file adds a healthcheck the README snippet omits
The compose example in the documentation defines the service, the container name, the `8080:8080` port mapping, the `chatclaw-data` volume at `/root/.config/chatclaw`, and `restart: unless-stopped`. The compose file in the repository has all of that plus a healthcheck block that the documented snippet leaves out:
healthcheck:
test: ["CMD-SHELL", "curl -sf http://localhost:8080/ || exit 1"]
interval: 30s
timeout: 5s
retries: 3
start_period: 10sTwo things follow. If you copy the documented file rather than the repository one, an orchestrator will treat the container as running even if the web interface is failing to serve, so you lose the restart signal that `restart: unless-stopped` would otherwise provide. And the check itself is a shell command using curl against the root path, which means the image is expected to contain curl for the healthcheck to report anything but failure.
The rest of the repository layout follows the same split between documentation and reality. There is a top-level `docker-compose.yml`, a `Taskfile.yml` for task automation, an `UpdateLog.md` for the full changelog, and `docs/readmes/` holding fourteen translated versions of this page, which is unusual attention to localisation for a project at version 0.9.7.
A file upload feature shipped in the same release as an OpenClaw downgrade
The changelog documents the tradeoff directly. On 2026-05-11 the project added what it calls a major chat file upload, with multiple file types and size limits and message sending updated to carry attachments. In the same entry, the OpenClaw version in the runtime configuration was downgraded to 2026.3.24 to maintain compatibility, and a `start-chatclaw.ps1` launch script was removed as no longer needed.
Reading those two items together is the useful part. The feature was added and the underlying component was moved backwards in the same release, which is what compatibility maintenance looks like when the upstream dependency is moving faster than the integration.
Two earlier entries give the shape of the maintenance work. On 2026-05-09 an OpenClaw factory reset was added to runtime settings, with a confirmation dialog and localisation in English and Chinese, and the frontend testing setup gained Playwright CDP debugging with a Dockerfile reworked for cross-compilation. On 2026-05-13 the release was localisation updates plus new error messages for plugin installation failures in English, Spanish, and French.
Plugin installation is the operation most likely to fail in the field, since skills come from a market, so translating exactly those error strings is a reasonable place to spend translation effort.
Agents are configured per role, with their own memory, skills, and reply style
The multi-agent mode is a configuration surface rather than a routing engine. You create several independent agents, each with its own role, its own memory, and its own skills, then assign identities such as a customer service specialist, a data analyst, or a copywriting assistant. Skills, knowledge base, and reply style are configured separately for each agent, and you switch between them in the interface.
The separation is what makes the mode useful: an agent with a narrow role and its own knowledge base produces answers scoped to that role, and switching is a deliberate act rather than something the assistant decides for you. Nothing in the description suggests the agents debate or hand work to each other, so treat this as configured specialisation rather than a multi-agent framework.
The skill market is the other half of that surface, with more than 5000 skills described as ready to use across office productivity, developer tools, multimedia creation, and smart home control, installable from within the app without writing code. The instructions also describe telling the bot to find already installed functions on your machine or install new extension plugins, which is the least conventional capability in the list.
Model choice is separate from agent choice. Providers listed include OpenAI, Claude, Gemini, Ollama, DeepSeek, Doubao, Qwen, Zhipu, and Grok, and there is a one-click authorisation that connects a ChatWiki account and syncs its credits, alongside custom model configuration.
Knowledge base ingestion is a fixed document list into sqlite-vec
The local knowledge base is described as an open source, ima-like manager. You upload documents, the system parses and splits them and converts them into vector embeddings, and they are stored in your private knowledge base for retrieval by the model. Organising by folder is supported.
The accepted types are named rather than open-ended: TXT, PDF, Word, Excel, CSV, HTML, and Markdown. That list is a good proxy for what the parser stack actually covers, because the Go module pins the matching Eino components: an HTML document parser, a markdown splitter, a recursive character splitter, and a semantic splitter.
Storage is SQLite with sqlite-vec for vector search, and the module pins `sqlite-vec-go-bindings` directly rather than shelling out, so vector search runs inside the same database file as the rest of the application state.
One extra dependency is easy to miss and explains a lot: `go-ego/gse` is a Chinese word segmenter, which is what lets the recursive and semantic splitters break mixed Chinese and English text into sensible chunks. For English-only documents that library is dead weight, and for Chinese documents it is load bearing.
The retrieval side is served by the Eino framework from CloudWeGo, the ByteDance project, along with tool components for MCP, DuckDuckGo, Wikipedia, and sequential thinking.
Scheduled tasks pair a graphical scheduler with cron expressions
Automation is configured two ways at once: a graphical scheduler for the common cases, and cron expressions when you need them. The frequencies named are every five minutes, hourly, and daily at a fixed time, and the stated use is scraping a specific page or data source, comparing the result to detect change, and watching for policy announcements, notice updates, or metric thresholds.
Alerts go out through the message channels rather than through the application window, so a task that fires while you are away reaches you on the phone. That is the same channel layer that carries the chat interface, and it is why the channel list is a core feature rather than an integration detail.
The channel list is broad: WeChat, DingTalk, WeCom, QQ, Feishu, and WhatsApp. Results are pushed automatically when processing finishes, without a manual refresh, and commands can be sent in the chat window to drive the agent remotely.
Two smaller interface features round out the same idea. Selecting any text on screen copies it into a floating quick-ask box, and a sidebar assistant can be docked beside another application so you can switch between differently configured assistants and send a reply into your current conversation. A multi-expert mode answers one question with several assistants side by side for comparison.
Editorial conclusion
ChatClaw fits a Windows or macOS user who wants an OpenClaw style agent with a document knowledge base and no scripting, especially if the work arrives over WeChat, DingTalk, Feishu, QQ, or WhatsApp. It does not fit a headless Linux deployment without reading the networking defaults, because server mode listens on all interfaces unless you override it. Before exposing it, set `WAILS_SERVER_HOST=127.0.0.1` or put it behind a proxy, confirm the binary and image you pull actually come from the repository you cloned, and note that the last commit is dated 2026-05-13 with release v0.9.7 on the same day, so newer skill or model behaviour may not be reflected in what you install.
Frequently asked questions
How do I run ChatClaw in server mode?
Download the Linux amd64 or arm64 server binary from the project's GitHub Releases, run chmod +x on it, and start it, then open http://localhost:8080. The service listens on 0.0.0.0:8080 by default and the bind address and port can be changed with WAILS_SERVER_HOST and WAILS_SERVER_PORT.
Can ChatClaw run without the desktop GUI?
Yes. Server mode runs headless and is reached through a browser, either from the published binaries or from the container image registry.cn-hangzhou.aliyuncs.com/chatwiki/chatclaw. Data persists in a named volume mounted at /root/.config/chatclaw.
Which document types can go into the ChatClaw knowledge base?
TXT, PDF, Word, Excel, CSV, HTML, and Markdown. Uploaded documents are parsed, split, converted into vector embeddings, and stored in the private local knowledge base, with folder based organisation, backed by SQLite with sqlite-vec for retrieval.
Which AI models does ChatClaw support?
The listed providers are OpenAI, Claude, Gemini, Ollama, DeepSeek, Doubao, Qwen, Zhipu, and Grok, and custom models can be configured. There is also a one-click authorisation that connects a ChatWiki account and synchronises its credits.
What messaging channels does ChatClaw connect to?
WeChat, DingTalk, WeCom, QQ, Feishu, and WhatsApp. Finished results are pushed to the channel automatically without a manual refresh, and you can send commands in the chat window to control the agent remotely. Scheduled tasks use the same channels for their alerts.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/zhimaai-chatclaw)Star history chart for your README
[](https://hysenlabs.com/projects/zhimaai-chatclaw)Was this useful?