M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
D

DeepChat

Community
Not a standard server
Local-first open-source AI agent desktop client unifying LLM providers, MCP, Skills, and ACP agents

This is not a typical connect-and-use MCP server; read the source repository before using it.

How to install ↓ Compare GitHub source repository ↗
Category
Other
Stars
★ 6.4k Very popular
Transport
stdio (local process) · SSE · Streamable HTTP
Runtime
Prebuilt binary
Credentials
No credential needed
License
Apache-2.0
Last commit
52FMRS · D

DeepChat is an actively maintained, popular open-source desktop AI assistant whose core value is a local-first, multi-model client with strong MCP-client capabilities (Resources/Prompts/Tools across stdio, SSE, and StreamableHTTP). It is not itself an MCP server that other clients can add, and no cached manifest was available, so this profile lists no specific tool identifiers — evaluate it as an AI agent desktop client rather than as a reusable tool server.

Strongest · Maintenance 15/20 Weakest · Security and permissions 6/20

Reliability
7/20
Security and permissions
6/20
Maintenance
15/20
Documentation
13/20
Setup experience
11/20
Why each score
Reliability 7/20
No server manifest was cached, so this review relies solely on README prose with no way to confirm that declared tools/resources/prompts match actual runtime behavior or that the MCP init/tool-list handshake works as described. The README mentions GitHub Action packaging for Windows/Linux builds but supplies no CI workflow files or committed test suite as verifiable execution evidence. Per the static calibration rule, reliability is capped at 12 without verifiable CI+test evidence, and given the total absence of manifest data it is scored lower at 7.
Security and permissions 6/20
This is a local-first Electron desktop client that bundles a 'code execution' inMemory MCP service and supports 'remote control' via Telegram, Feishu/Lark, QQBot, Discord, and WeChat iLink — including remotely answering pending permission prompts, stopping runs, and opening desktop sessions. This suggests local confirmation dialogs could be bypassed by a bound remote bot, which is a real overreach/least-privilege concern. The README only offers marketing-level claims ('Security-First Design', 'reserved encryption interfaces') without concrete detail on permission scoping, data boundaries, or the strength of the /pair authentication mechanism for remote binding. No evidence of credential theft or fabricated results was found, so this does not meet the 0-4 red-line bar, but the disclosed risk surface is real and underspecified, warranting a score of 6.
Maintenance 15/20
The project uses the Apache-2.0 license, has 6200 stars, only 9 open issues, ships through multiple channels (GitHub Releases, official website, macOS Homebrew cask), and references contribution guidelines, a wiki, and multilingual docs — signs of an actively governed project. However, the supplied material contains no explicit recent-commit or release-date evidence to confirm update cadence, so it falls short of full marks, scored 15.
Documentation 13/20
The README is well-structured with a table of contents and sections covering features, Tape/Trace, Skills, ACP, and Remote Control, linking out to docs/README.md and a developer guide. However, as the object under review is nominally an 'MCP server,' the documentation lacks concrete tool parameter specs, limits, cost information, or a troubleshooting section, and the missing manifest prevents cross-checking documented behavior against actual tool implementation, scored 13.
Setup experience 11/20
Three install paths are offered (GitHub Releases binaries, official website, macOS Homebrew), plus development setup steps (pnpm install, runtime install, a documented Windows symlink permission caveat) and per-platform build commands — reasonably clear for an Electron app. Per the static calibration rule, setup is capped at 15 without verifiable CI+test execution evidence; additionally, this project is itself an MCP client/host rather than a typical MCP server with a one-step client config example for mainstream MCP clients, so it is scored 11.

Static review · not runListed 2026-08-06

Read the FMRS scoring method →

Fit and risk

Best for

  • Users who want a single desktop app that unifies many cloud LLM providers (OpenAI, Anthropic, Gemini, DeepSeek, etc.) and local Ollama models
  • Users who need to connect to multiple external MCP tool services with local-first storage and inspectable session history (Tape/Trace)
  • Users who want to reuse Agent Skills built for Claude Code, Cursor, Windsurf, and similar tools
  • Developers who want coding-focused ACP agents integrated into the same chat-style model selector

Not for

  • Users looking for an MCP server to add into Claude Desktop or another AI client's config (DeepChat is an MCP client/host, not a server to be added elsewhere)
  • Users who want a lightweight CLI-only tool rather than a full Electron desktop application
  • Users needing a native mobile app (currently Windows/macOS/Linux desktop only; mobile access is indirect, via remote control)

Required permissions

  • Local file system read/write access for session Tape storage, Skills files (imported from folders/ZIP/URL), and app configuration
  • Ability to spawn local subprocesses, used to run stdio-based MCP services and the bundled Node.js runtime
  • Network access to call cloud LLM APIs, search services (e.g. Brave Search), and SSE/StreamableHTTP MCP services
  • Remote-control bot credentials (Telegram/Feishu/QQBot/Discord/WeChat iLink tokens) used to bind and remotely operate desktop sessions
  • Local Ollama management access to download, deploy, and run local models

Risks and side effects

  • Third-party MCP services connected via stdio run as local processes under the user's permissions — only install MCP services from trusted sources
  • Imported Skills may include optional scripts that execute with the app's permissions once enabled — vet the source before enabling
  • If a remote-control bot token is leaked, an attacker could remotely control the local session, switch models, or read pending interactions via the messaging app
  • Some search functionality works by simulating user browsing of search engines, which may raise terms-of-service considerations for target sites
  • The README describes encryption/obfuscation interfaces as "reserved" capabilities — verify what is actually enabled in a given release before relying on it

Setup

Before you start

Runtime：Prebuilt binary

1) Download the DeepChat desktop app from GitHub Releases, the official website, or (macOS only) brew install --cask deepchat; 2) Open Settings → Model Providers and add your cloud LLM API keys, or configure local Ollama; 3) To extend MCP capabilities, add/manage MCP services in Settings (stdio, SSE, or StreamableHTTP transports, or one-click DeepLink install) — the bundled Node.js runtime can be installed via pnpm run installRuntime; 4) For Skills, import them under Settings → Skills and enable per conversation; for ACP agents, enable built-in or custom commands under Settings → ACP Agents; 5) Click "+" to start a new conversation and select a model.

Check that it works

Launch the DeepChat desktop app and open the MCP services list; if an added service shows its Resources/Prompts/Tools, the setup works. Note: this entry is an MCP client (desktop app), not a server, so there is nothing to configure in a client.

Troubleshooting

  1. Features not working after install: confirm `pnpm run installRuntime` was run to install the bundled Node.js runtime; if you see `No module named 'distutils'`, run `pip install setuptools` first
  2. `pnpm` operations failing on Windows: enable Developer Mode in Settings or use an administrator account so symlinks/hardlinks can be created
  3. Model not responding: check that API keys are correctly set under Model Providers, or that the local Ollama service is running
  4. MCP service connection failing: verify the transport used (stdio/SSE/StreamableHTTP) and inspect parameters/return data in the tool-call debug view
  5. ACP agent won't start: check that the built-in agent is enabled under Settings → ACP Agents, or that a custom command path is correct
  6. Remote control unresponsive: confirm the messaging platform is bound under Settings → Remote and that the bot token is still valid

Things to try

Once connected, you can ask your AI assistant things like:

  • Install an MCP service in DeepChat settings via DeepLink and enable its tools
  • Add a Stdio MCP server and inspect the tool call parameters and return data
  • Enable the built-in inMemory services to run code or fetch web information

Use cases

Daily assistant: answering questions, giving suggestions, assisting with writing and creation
Development aid: code generation, debugging, technical problem solving, and running ACP coding agents
Learning tool: concept explanation, knowledge exploration, learning guidance
Content creation: copywriting, creative inspiration, content optimization
Data analysis: data interpretation, chart generation, report writing
Remote operation: managing a running desktop session from a phone via Telegram, Discord, Feishu/Lark, etc.

Overview

DeepChat is ThinkInAIXYZ's open-source, local-first AI agent desktop client (Apache-2.0) for Windows, macOS, and Linux. Note that DeepChat is not itself an MCP server that other clients connect to — it is an MCP client/host application: it connects to external MCP services over stdio, SSE, and StreamableHTTP transports (with full Resources/Prompts/Tools support), ships a built-in Node.js runtime so npx/node-based MCP services work out of the box, and provides its own inMemory services for code execution, web information retrieval, and file operations. Beyond MCP, DeepChat supports installing and enabling Agent-Skills-compatible Skills per conversation, running ACP (Agent Client Protocol) agents as model-selector entries, and remote-controlling sessions from Telegram, Feishu/Lark, QQBot, Discord, and WeChat iLink. The README does not include a config example for adding DeepChat as an MCP server to another client, and no server manifest was available, so no specific tool identifiers can be listed here.

Similar servers

Telegram MCP Server 70 · B

Telegram MCP server powered by Telethon to let MCP clients read chats, manage groups, and send/modify messages, media, contacts, and settings.

★ 2.0k · Tools 76 Compare with this →
Compare with other MCP servers

Source revision 145e51763262 Data synced 2026-11-04 Read the FMRS scoring method