Configuration you should tune

1. The files most worth tuning
2. config.json shape
3. Adding your own MCP servers
4. Tools are global, configuration is per-project

---

Everything is driven by .aiflow/config.json (committed, no secrets). Edit it interactively with aiflow change-settings (re-renders .mcp.json, hooks, branching, memory) — switch version control (git/svn), pick different Ollama models, or turn token saving off entirely with aiflow change-settings --no-token-saving. Secrets always stay in .env (gitignored, never global).

The files most worth tuning

• AGENTS.md — the operating rules any coding agent reads (Claude Code, GitHub Copilot, OpenAI Codex CLI): project overview, architecture hints, code style, task workflow, git rules, the memory/context stack, communication. Fill the [EDIT ME] blocks (§1 overview, §2 architecture) — this is the single biggest quality lever. CLAUDE.md is just a one-line @AGENTS.md import for Claude Code — edit AGENTS.md, not CLAUDE.md. See Multi-Agent Support.
• .aiflow/team-prefs.json (the “preferences” file) — shared, versioned team/user preferences: code style preset, language, conventions. Committed so the team inherits them; overrides AGENTS.md §3.
• .claude/memory/ — project-aim.md (goal + architecture), dev-environment.md, memory-policy.md (the retrieval routing + learning intensity). Claude Code only — read these files directly if your agent doesn’t support persistent memory. Keep them current.
• .claude/settings.json — permissions (allow/deny), hooks (caveman, formatter, beads-sync), MCP allow-list. Claude Code only.
• .aiflow/branching.json / docs/branching.md — the branching + release model.
• .env — all tokens/keys.

config.json shape

{
  "caveman":  { "enabled": true, "mode": "full" },
  "rtk":      { "enabled": true },
  "router":   { "enabled": false },
  "graphify": { "enabled": true },
  "taskmaster": { "enabled": true },
  "mcp":      { "filesystem": true, "context7": true, "cocoindex": true },
  "memory":   { "enabled": true, "graph": true, "intensity": "aggressive" },
  "agents":   { "claude": true, "copilot": false, "codex": false },
  "claude":   { "auth": "apikey" },
  "vcs":      { "system": "git" },
  "remote":   { "type": "github", "baseUrl": "https://github.com",
                "api": "github-api", "tokenEnv": "GITHUB_TOKEN", "mcp": "github" },
  "gitkraken":{ "enabled": false },
  "codexsaver":{ "enabled": false, "provider": "deepseek", "apiKeyEnv": "DEEPSEEK_API_KEY" },
  "sync":     { "askOnClose": true, "pullOnStart": true },
  "ollama":   { "enabled": false, "url": "http://localhost:11434", "models": [] },
  "teamPrefs":{ "enabled": false, "codeStyle": "google" },
  "project":  { "aim": "…", "architecture": "…" },
  "dev":      { "os": "windows", "ide": "vscode" },
  "git":      { "model": "gitflow", "strict": true, "prOnly": true,
                "autoRelease": true, "versionStrategy": "semver", "releaseTags": true, "chore": true }
}

Adding your own MCP servers

aiflow generates .mcp.json from the config, but you can add any extra MCP server — your edits to servers aiflow doesn’t manage are preserved on re-render:

{
  "mcpServers": {
    "my-server": {
      "command": "npx",
      "args": ["-y", "@scope/my-mcp-server"],
      "env": { "MY_TOKEN": "${MY_TOKEN}" }   // secrets via .env, never inline
    }
  }
}

Then allow it in .claude/settings.json under permissions.allow (e.g. "mcp__my-server") and put any secret in .env. For community-vetted servers, browse npx claude-code-templates@latest. Tip: prefer a focused MCP over a broad one — fewer tools = less context and fewer wrong turns.

Tools are global, configuration is per-project

• Tools / binaries — installed once per user (npm -g, uv tool, brew/winget); shared across projects. aiflow install-deps puts them there; the router config lives in your home dir.
• Configuration & secrets — per project: .env, .aiflow/config.json, AGENTS.md, .mcp.json, .claude/, .githooks/, memory. Switching projects switches config; nothing leaks between them.