AI Agents and Skills

AI agents & skills

Agents tab

StackRadar detects an agent when its CLI is on your PATH or its config folder exists:

Agent	CLI	Folder(s) checked	Extras read
Claude Code	claude	~/.claude	sessions, tokens, models, MCP servers (~/.claude.json), hooks, enabled plugins
Codex CLI	codex	~/.codex	sessions, token totals, MCP servers (config.toml)
Hermes Agent (Nous Research)	hermes	~/.hermes	sessions, MCP servers (config.yaml), cron jobs
OpenClaw (ex-Clawdbot / Moltbot)	openclaw	~/.openclaw, ~/.clawdbot, ~/.moltbot	sessions, MCP, cron jobs, gateway port 18789
Paperclip	paperclipai	~/.paperclip	server port 3100
Gemini CLI	gemini	~/.gemini	sessions, MCP servers
Cursor · Windsurf · Copilot CLI · OpenCode · Goose · Aider · Qwen Code · Amp · Kiro · Continue · Cline/Roo	various	various	MCP servers where stored
Ollama · LM Studio	ollama · lms	~/.ollama · ~/.lmstudio	local models, ports 11434 / 1234

Each card shows version, sessions, last activity, skills, MCP server names (never their env values or keys), logged tokens, on-disk size, matching running processes and ports. A status badge says running, installed or config only. Agents StackRadar knows about but didn't find are listed at the bottom, so you know what was checked.

Tokens by agent, surface and model

Above the cards, a table splits logged tokens by model (in / out, replies, sessions) and by agent and where the session ran: CLI, IDE (VS Code, JetBrains), Desktop app or Web / cloud. Claude Code records this in each transcript (entrypoint); Codex in originator.

Drill into an agent

Click a card:

• Installed vs newest version (Check asks npm / PyPI / Homebrew) and ⬆ Update (Claude Code uses claude update).
• Sessions: title, first prompt, project, where it ran, model, tokens in / out, last active, size. Per session: 👁 read it, ⇢ hand off, 🏷 tag (use to delete to mark clean-up candidates), 🗄 archive (hidden from the agent, kept in ~/.stackradar/archived-sessions, restore any time) and 🗑 delete (to the Trash). Select several for bulk tag / archive / delete. Filter by tag.
• Tokens & models for this agent.
• Processes with start time, CPU and memory, Stop / Force each, and Stop all but the newest when an agent has left old processes behind.
• Its skills and MCP servers.

Continue with another agent

Hand a session, or a whole project, to another agent and model:

• From a session: ⇢ hand off. From a project's drawer: Continue with another agent (StackRadar marks projects that have agent sessions but no activity for 14+ days as stalled). Pick Claude Code, Codex, Gemini CLI, Aider or OpenCode and, optionally, a model (suggestions come from models you've used and your local Ollama models).
• StackRadar writes a brief to ~/.stackradar/handoffs/: the original request, later instructions, where the last agent stopped, open to-dos and the files it touched (or, with no session, what StackRadar knows about the project).
• You get the command (cd … && codex -m <model> "$(cat brief.md)", claude --resume <id> to continue in the same agent …) to copy, or Open in Terminal to start it right away.

Skills tab

What counts as a skill

• SKILL.md folders: ~/.claude/skills, Claude Code plugin skills, ~/.codex/skills, ~/.agents/skills, ~/.hermes/skills, ~/.openclaw/skills and ~/.openclaw/workspace/skills, ~/.gemini/skills, ~/.cursor/skills, OpenCode, Goose, Copilot …
• Slash commands (~/.claude/commands/*.md, ~/.codex/prompts, Gemini commands/*.toml) and sub-agents (~/.claude/agents/*.md).
• Project-level copies: <project>/.claude/skills, .agents/skills, .codex/skills, skills/*/SKILL.md …
• skills-lock.json pins.

How usage is counted

• Claude Code: every Skill tool call and every /slash-command in ~/.claude/projects/**/*.jsonl (newest first, up to 300 MB of logs), with the project (cwd) it happened in.
• Codex / Hermes / OpenClaw: mentions of <skill-folder>/SKILL.md in their session logs (they read the file instead of calling a tool). Best effort.

What you see

• Installed / used / never used / duplicate names / used-but-not-installed cards (click to filter).
• Per skill: agent, scope (global or project), usage bar, which agents used it, last used, which projects, size, 📂 reveal.
• Duplicate skills: the same name in several places, with every location, its agent, modified date, uses and a version tag (same number = identical). Identical compares every file in the skill folder, not just SKILL.md. Keep this, link the rest moves the copy you choose to the shared ~/.agents/skills/<name> and replaces every other location with a link to it (old copies go to the Trash), so all agents use one copy.
• Per skill: 🗑 delete (to the Trash; for a link, only the link is removed) and ⇢ use it in another agent (adds a link in that agent's skills folder). Plugin skills are managed by the agent's plugin command instead.
• Most-used Claude Code tools and used but not installed (built-in, plugin-provided or deleted skills).
• In each project's drawer: skills installed in that project and skills used while working there.