Where DeepSeek stores context files, settings, and MCP servers

DeepSeek is first a model provider, not a harness — and that shapes where its files live. There is no single “DeepSeek CLI” with one settled convention; instead there’s an official terminal agent (Deep Code), a handful of community tools, and a lot of DeepSeek usage that happens inside other tools whose file conventions apply instead. This page maps what verifiably exists. ~ is your home directory.

At a glance

Artifact	Global	Per-project	Format
Settings (Deep Code, official)	~/.deepcode/settings.json	—	JSON
Skills (Deep Code)	~/.agents/skills/<name>/SKILL.md	.deepcode/skills/<name>/SKILL.md	Markdown + YAML frontmatter
Config (DeepSeek-TUI, community)	~/.deepseek/config.toml	—	TOML
MCP servers (DeepSeek-TUI, community)	~/.deepseek/mcp.json	—	JSON
Context / rules	— no standard DEEPSEEK.md exists —	the harness’s own files (AGENTS.md, CLAUDE.md, …)	Markdown

There is no standard DEEPSEEK.md

Nothing official reads a DEEPSEEK.md

Unlike CLAUDE.md, GEMINI.md, or QWEN.md, no DeepSeek tool documents a DEEPSEEK.md context file. If a repo has one, it’s either aspirational or fed to a harness that lets you rename its context file (Qwen Code’s context.fileName, for example). Don’t assume an agent is reading it — verify which tool, if any, actually loads it.

The honest picture: DeepSeek models are most often driven through harnesses like Cline, GitHub Copilot CLI, OpenCode, or Codex-compatible CLIs pointed at the DeepSeek API. In those setups your context files are the harness’s — AGENTS.md, .clinerules, CLAUDE.md — documented on their own pages.

Deep Code: the official terminal agent

Deep Code is DeepSeek’s open-source terminal coding assistant for DeepSeek-V4, documented in the official DeepSeek API docs. Its configuration is one file:

~/.deepcode/settings.json

{
  "MODEL": "deepseek-chat",
  "BASE_URL": "https://api.deepseek.com",
  "API_KEY": "sk-...",
  "thinkingEnabled": true,
  "reasoningEffort": "medium"
}

The API_KEY sits directly in this file — it’s a credential store, not just configuration. It belongs in no backup you share and no index. The same file is shared with the Deep Code VS Code extension.

Deep Code skills — the gotcha

Deep Code supports Agent Skills in the familiar SKILL.md folder convention, but its two locations don’t mirror each other:

• User-level: ~/.agents/skills/<name>/SKILL.md
• Project-level: .deepcode/skills/<name>/SKILL.md

User skills live in ~/.agents/, not ~/.deepcode/

The global skills directory is ~/.agents/skills/ — a cross-tool location, not inside ~/.deepcode/. That’s good news for sharing skills between tools that adopt the same convention, and bad news for anyone backing up ~/.deepcode/ and assuming they got everything. Project skills, meanwhile, are under .deepcode/.

DeepSeek-TUI and other community tools

The community has filled the CLI gap several times over, each tool with its own layout — verify against the tool you actually installed:

• DeepSeek-TUI keeps its config at ~/.deepseek/config.toml (API key, default model) and reads MCP server definitions from ~/.deepseek/mcp.json. This is the one place a ~/.deepseek/ folder verifiably appears.
• Other deepseek-cli projects on npm and PyPI use their own paths (e.g. ~/.config/deepseek-cli/config.json).

None of these are official conventions, and none read a shared context file.

Using DeepSeek through other tools

If you drive DeepSeek through Claude Code-style CLIs, Codex-compatible tools, or Cursor, everything on those pages applies unchanged — the model behind the harness doesn’t move the files. The one DeepSeek-specific artifact is the API key in whichever config points BASE_URL at api.deepseek.com.

Seeing all of it at once

Quartermaster’s scanner covers the DeepSeek ecosystem as it actually is: Deep Code and community configs where they verifiably live, plus forward-compatible patterns (DEEPSEEK.md, .deepseek/settings.json) so anything that adopts those names is indexed the moment it appears — alongside Claude Code, Cursor, Codex, and 8 more tools. Secrets are never indexed — env/header key names kept, values dropped. See the full discovery map, how conversion moves artifacts between tools, or the quickstart.