Validity and method: ZeroChat's data was checked in its code (version 9.11.0, commit 170ee03). The others' data comes from their official public documentation: the new rows (remote access and terminal runs) were consulted on 24 October 2026 and the rest between 4 and 16 October. What that documentation does not state is shown as "not stated", not as "no". Memory and performance were not measured.
1. What is compared
ZeroChat aims to be the single frontend for models served by their own engine (vLLM, llama.cpp, LM Studio, SGLang or Ollama), switching models by task. It is therefore compared with the interfaces usually placed in front of a local server, not with those aimed only at cloud providers.
Since the 9.2.3 review the comparison adds two dimensions where ZeroChat has changed: remote access to the working environment and running a prompt from the terminal.
2. Matrix
| Dimension | ZeroChat | Open WebUI | LM Studio | llama.cpp interface | Jan |
|---|---|---|---|---|---|
| Deployment | Static web app and an optional single-file Python local server | Server (Docker or pip) with a database | Desktop app with its own engine | Bundled with llama-server | Desktop app with embedded llama.cpp |
| Connect to a running server | Yes: OpenAI compatible, Ollama and cloud APIs | Yes | Focused on its own engine; configurable | Only its own server | Yes, as a remote provider |
| Separate and forwarded reasoning | Yes, in a collapsible block and forwarded between agent steps | Yes | Yes | Yes | Not stated |
| MCP | stdio services managed by the local server; remote HTTP ones through mcp-remote, with OAuth | Native (Streamable HTTP) | Yes, local and remote | Not stated | Yes |
| Model browsing | browser_action or Playwright MCP with its own engines or the user's browser | Through tools or MCP | Through MCP | Not stated | Through MCP |
| Tool permissions | Global policies, rules per directory and command, a signature per call | By administrator and roles | Interactive confirmation with editable arguments | Not stated | By configuration |
| Computer files and commands | Yes, through the local server, without containers, plus a diff viewer | Through tools or MCP | Through MCP | Not stated | Through MCP |
| Documents | Word, PowerPoint, Excel, OpenDocument, .doc, .xls and PDF, in the chat and in RAG | Through server-side extraction | Document attachments | Not stated | Basic documents |
| Vision and web images | Yes, with web image downloads for the model | Yes, image attachments | Yes, with vision models | Yes, with multimodal support | Yes, with vision models |
| Subagents | Yes, read-only (delegate_task) | Not stated | Not stated | Not stated | Not stated |
| RAG | Lexical in the browser (Orama in IndexedDB) | Vector with a database | Document attachments | Not stated | Local retrieval |
| Request inspection and editing | Editable interceptor before each send | Server logs | Developer logs | Console logs | Limited |
| Remote access | The full environment (agent, files and commands) from any browser, by SSH tunnel or reverse proxy, with sign-in through the system account | Web server with its own user accounts | The model API on the local network or through LM Link (preview); the interface is a desktop app | The llama-server one; its API key protects the API | The model API on the local network, with an API key; the interface is a desktop app |
| Prompt from the terminal | zerochat --prompt: the full agent with tools and profile, answer on stdout and conversation export | Not stated | lms chat -p: the model's answer on stdout | llama-cli, separate from the interface | Not stated |
| Multi-user | No: one system account | Yes, with roles and authentication | No | No | No |
3. Analysis
Versus the llama.cpp interface
It is the minimal option: it ships with the server and needs nothing else. It is good for trying a model, but it has no agent, computer tools, permissions or MCP. ZeroChat connects to the same llama-server and adds the whole working environment.
Versus LM Studio
LM Studio combines engine and desktop interface, supports MCP with a confirmation for each call, and its CLI answers a prompt from the terminal. Its remote access is that of the model API. ZeroChat does not depend on the engine and brings the agent with the computer's files and commands to the remote browser; its --prompt runs that full agent, not just the model.
Versus Open WebUI
Open WebUI is a client-server suite with a database, user accounts and vector RAG, designed for teams. ZeroChat is a static application without a server database, for one person, that starts without Docker, keeps data in the browser and also runs on Android with Termux.
Versus Jan
A desktop application for running GGUF models on the computer itself, with a local API that can be opened to the network. For people who already serve their models with vLLM or llama.cpp, ZeroChat offers more control over the connection, transparent requests and built-in technical tools.
4. What sets ZeroChat apart for this use
- Engine independence: one client for vLLM, llama.cpp, LM Studio, SGLang or Ollama.
- The working environment remotely: agent, files and commands of the computer from a phone or another computer, without exposing the local server and with the system account as sign-in.
- The agent from the terminal:
zerochat --promptruns a complete task with tools and returns only the answer, suitable for scripts and cron. - Profiles that prepare the environment: each profile chooses the model, MCP services and browser.
- Reasoning preserved between steps and native documents (office and PDF) without heavy server dependencies.
- Editable interceptor for the exact JSON sent to the model.
Accepted trade-offs: lexical-only RAG, a single person, and data in the browser that should be exported.
5. When to choose each one
- ZeroChat: one person who wants an agent for programming or technical work on a local model served separately, with controlled access to the computer, also away from home.
- LM Studio: a graphical application that downloads, runs and serves the model.
- llama.cpp interface: quick checks without installing anything else.
- Open WebUI: several users, central administration or vector RAG.
- Jan: simple personal use of GGUF models on the desktop.
6. Sources
- Open WebUI: MCP (consulted in October 2026)
- Open WebUI: llama.cpp integration (consulted in October 2026)
- LM Studio: MCP and release notes (consulted in October 2026)
- LM Studio:
lms chat, serve on the local network and LM Link (consulted on 24 October 2026) - Jan: changelog and local API server (consulted on 24 October 2026)
- llama.cpp:
llama-server(consulted on 24 October 2026)
Authorship
- Review and writing: Claude Opus 5.5 in Claude Code, from the dated sources and ZeroChat's code at commit
170ee03.