Adeptly turns a sentence into a full plan — with the right Claude Code features baked into every section. Then it runs the plan, scores how lean it is on tokens, and shows what you spent. All on your machine. No API key, no backend, no telemetry.
Plan Mode, subagents, skills, hooks, MCP, auto-memory — most people never find them. And when a few devs share a repo, nobody knows the plan until the PR lands. Adeptly closes both gaps.
Adeptly names the exact features your plan should use — inline, with hover explanations — so you learn them by reading your own plans. A live Coverage score makes it measurable.
Adeptly is a thin GUI over your local claude CLI. Plans live as markdown in git — that's the whole team-sync model. Nothing new leaves your machine.
One plan-first workflow — and you can stop, hand off, or take over at any step.
Adeptly writes the plan and underlines the exact Claude Code features to use — right in the sections they apply to. Hover any one for what it is and why it's here. Nobody else explains Claude Code like this.
Start in Plan Mode before writing code — the change spans sessions and middleware.
Map the current flow with an Explore subagent, then implement behind a flag.
Run /security-review before merging the session-token changes.
Two scores over your plan: how many Claude Code features it actually uses, and how many money-saving habits it follows. Each gap is one click to bake into the plan.
Every call Adeptly makes records its real token cost and cache-hit. Flip to Whole account to see machine-wide spend parsed from ~/.claude — tokens exact, cost estimated. 100% local, no telemetry.
Fifteen more reasons it stays open in a tab all day.
Split a plan across parallel sessions. Each brief carries the same Working Agreement and its own file lane — so two Claude tabs don't collide or skip the safety checks.
Runs an approved plan as roles — Architect → Builder → Medic → Reviewer + Security (parallel) → PR. Dry-run by default; live is double-gated.
Leave a resume note one Claude session hands to the next — the bridge sessions can't make live.
Ask your own self-updating claude what features are new and fold them in. No backend, no doc-scraping.
Detects your stack from manifests so every recommendation, path, and command fits what you actually build with.
Browse every past crew run — mode, pass count, cost, PR — and re-run any of them in one click.
Plans and approval state are files. Every dev sees the same thing; git is the only sync layer. No accounts.
Dictate in chat (push-to-talk) or paste a screenshot and ask Claude to read it.
--host for LAN, --tunnel for a Cloudflare quick-tunnel — with a loud, honest trust warning.
Structured workflow for a plan: subagents by purpose, skills with when-to-invoke, hooks, expected turns & cost.
⌘K to search and open plans, ⌘I for chat, focus mode to hide the panels.
The plans list shows git author + dirty-state, so you see who last touched what at a glance.
One click turns plan + recipe into a paste-ready Claude Code prompt for a fresh session.
Drag-resize splitters, a mint-green pulse when a plan gets approved, and both themes done right.
Every crew run is written to runs/<slug>/ as JSON + an append-only audit.jsonl.
Cursor, Continue, and Aider generate code. Adeptly starts one step upstream — the plan that becomes a great prompt — and closes the loop with a crew that runs it.
| Capability | Adeptly | Raw Claude Code | Code-gen IDEs |
|---|---|---|---|
| Plan-first, approve before code | ✓ | manual | — |
| Teaches every Claude Code feature inline | ✓ | — | — |
| Coverage + token-hygiene score | ✓ | — | — |
| Shows your real token spend | ✓ | — | — |
| Runs the plan as a role crew → PR | ✓ | DIY | partial |
| Team sync with zero backend (git) | ✓ | — | — |
| Runs locally, no API key of its own | ✓ | ✓ | varies |
No SaaS backend, no API key, no telemetry, no account. If you trust Claude Code, Adeptly adds zero new trust surface.
| What | Where it runs | What it sends |
|---|---|---|
| Adeptly web UI | localhost | nothing |
| Adeptly node server | your box | nothing |
| claude --print subprocess | your box | your prompt → your CLI → Anthropic, exactly as if you typed it |
Pulled live from the npm registry and GitHub right now — no marketing math, just the real counters.
Needs Node 18.17+ and a signed-in Claude Code CLI.
npm i -g @anthropic-ai/claude-code && npx adeptly