ChatLLM Web v3.2.0
v3.2 turns the browser-native Local Model Studio into a Local Agent Workspace with native WebLLM function calling, observable tool runs, and approval-gated artifacts.
Highlights
- Per-conversation Chat and Agent modes.
- Native tool calling with Hermes 2 Pro Mistral 7B, Hermes 2 Pro Llama 3 8B, and Hermes 3 Llama 3.1 8B.
- Eight sandboxed local tools for selected context, arithmetic, and conversation artifacts.
- Automatic read-only tool execution and explicit diff approval for every artifact write.
- Beautiful UI Agent Run Rail, Tool Chips, Task Rows, Diff Table, and Artifact panel.
- Eight-step run limit with stop, retained progress, error attribution, and retry.
- Markdown, code, JSON, and text artifacts stored in IndexedDB and available for copy or download.
- 20 curated models and 65 logical models backed by 163 WebLLM 0.2.84 records.
Security boundary
- File tools can access only the files selected for the current run.
- Artifact tools can access only the current conversation.
- Arithmetic uses a dedicated parser without dynamic code execution.
- Artifact creation and replacement pause until the user approves the displayed line diff.
- Shell, direct disk writes, network search, MCP, RAG, and third-party data sources are outside the v3.2 runtime.
Data and upgrade
v3.2 starts with the new chatllm-v3.2 IndexedDB schema. Existing v3.1 browser data remains untouched and is not loaded into v3.2. Model assets already managed by the WebLLM Cache API remain provider-controlled browser cache entries.
Links
- Live: https://chatllm-web.pages.dev
- Repository: https://github.com/Ryan-yang125/ChatLLM-Web
- Models: https://chatllm-web.pages.dev/models