agentpeek

A TUI for inspecting agent CLI configurations: what's installed, what's enabled, what's broken.

agentpeek scans your Claude Code config (~/.claude/ and any project-level .claude/ directory it finds by walking up from cwd) and surfaces every loose file the agent depends on: settings, hooks, slash commands, memory files, MCP servers, user/project subagents, path-scoped rules, installed plugins (with their full contents: manifest, skills, agents, commands, hooks, MCPs), and any health warnings detected during the scan.

Install

pipx install agentpeek

Then:

agentpeek                      # scan ~/.claude and the nearest project .claude
agentpeek --root /path/to/dir  # scan a specific directory
agentpeek --version

Requires Python ≥ 3.11. Linux and macOS only.

What you get

• Three-zone TUI: sidebar of 11 categories on the left, items list in the middle, detail pane on the right. Drill-down on every category.
• Multi-scope: user-scoped (~/.claude/) and project-scoped (.claude/ found by walking up from cwd) configurations shown side by side with [U] / [P] badges. Health checks compare the two and flag conflicts.
• Subagents: user and project agents under agents/*.md are scanned recursively, with full frontmatter (tools, model, permission mode, memory scope, color, and the rest) shown in the detail card.
• Path-scoped rules: .claude/rules/*.md files at user and project scope are listed with their paths: globs (or marked "always loaded" when the field is absent), so you can answer "why is Claude doing X here?" by category instead of by guessing.
• Plugin contents: every installed plugin is opened up: manifest, skills, agents, slash commands, hooks, MCP servers, and the marketplace's auto-update state. Plugin-contributed commands / hooks / MCPs also merge into the top-level categories with a [plug:<id>] provenance tag.
• Memory layer: CLAUDE.md, CLAUDE.local.md (project-root personal overrides), and subagent persistent memory directories (agent-memory/ and agent-memory-local/) are all surfaced. CLAUDE.md @path/to/file imports are resolved recursively (up to 5 hops, cycles detected) so you see what Claude actually loads.
• Auto-memory: Claude Code project memories under ~/.claude/projects/<encoded>/memory/ are scanned and labelled with their actual project path (resolved from session logs, not the lossy directory encoding).
• ~/.claude.json: user-scope MCP servers, per-project trust state (allowed tools, enabled / disabled .mcp.json servers), and OAuth session presence (display-only - the token value is never read or shown). Folded into the MCP and Settings cards.
• Health checks: orphan hook scripts, missing plugin install paths, conflicting key bindings, scope override conflicts, cross-scope layered memory, hooks referencing dynamic env vars (${CLAUDE_PROJECT_DIR}, $HOME, ...) flagged so you know which won't resolve at scan time, and more. Severity-coloured warning cards in the detail pane.
• Write actions: update, enable/disable, uninstall a plugin or refresh a marketplace without leaving the TUI. Calls shell out to claude plugin ...; agentpeek never edits Claude state files directly. Re-scan is automatic after every action. Pass --read-only to disable.

Keybindings

The filter shows (filtered: N / M) in the title while active. Press Enter inside the filter to keep the filter and jump back to the items list; press Esc to clear.

Theming

Built on Textual: the theme follows Ctrl+P (palette) and reflows colours across cards, borders, badges, and severity indicators. Default theme is textual-dark.

Supported agent CLIs

• Claude Code: full support today.
• Support for other agents (Codex, Gemini CLI, Cursor) coming.

License

GPL-3.0-only. See LICENSE.