{"id":51616520,"url":"https://github.com/neural-llm-io/power-claude","last_synced_at":"2026-07-28T03:00:18.892Z","repository":{"id":357950936,"uuid":"1239277817","full_name":"neural-llm-io/power-claude","owner":"neural-llm-io","description":"Local orchestration for Claude Code — session recovery, scheduling, worktrees, Token Saver, and dashboards via official CLI transport. Independent third-party software.","archived":false,"fork":false,"pushed_at":"2026-07-11T22:27:53.000Z","size":9461,"stargazers_count":4,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-07-12T14:14:42.114Z","etag":null,"topics":["ai","ai-coding","anthropic","claude","claude-code","cursor","developer-tools","llm","local-orchestration","open-vsx","productivity","session-recovery","token-saver","vscode-extension","windsurf"],"latest_commit_sha":null,"homepage":"https://neural-llm.com/power-claude","language":null,"has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"other","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/neural-llm-io.png","metadata":{"files":{"readme":"README.md","changelog":"CHANGELOG.md","contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-05-15T00:02:23.000Z","updated_at":"2026-07-10T15:46:53.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/neural-llm-io/power-claude","commit_stats":null,"previous_names":["neural-llm-io/power-claude"],"tags_count":14,"template":false,"template_full_name":null,"purl":"pkg:github/neural-llm-io/power-claude","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/neural-llm-io%2Fpower-claude","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/neural-llm-io%2Fpower-claude/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/neural-llm-io%2Fpower-claude/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/neural-llm-io%2Fpower-claude/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/neural-llm-io","download_url":"https://codeload.github.com/neural-llm-io/power-claude/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/neural-llm-io%2Fpower-claude/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":35973263,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-07-20T02:08:10.276Z","status":"online","status_checked_at":"2026-07-28T02:00:06.341Z","response_time":109,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["ai","ai-coding","anthropic","claude","claude-code","cursor","developer-tools","llm","local-orchestration","open-vsx","productivity","session-recovery","token-saver","vscode-extension","windsurf"],"created_at":"2026-07-12T14:00:37.474Z","updated_at":"2026-07-28T03:00:18.832Z","avatar_url":"https://github.com/neural-llm-io.png","language":null,"funding_links":[],"categories":[],"sub_categories":[],"readme":"# Power Claude — Spend less. Run faster. Never stall on Claude Code.\n\n\u003e ### 🎉 Due to popular demand — now on the VS Code Marketplace, Open VSX \u0026 npm\n\u003e Power Claude has been available **direct from [neural-llm.com](https://neural-llm.com) since December 2025**. By popular request, it's now one click away in your editor's marketplace too — the same battle-tested tool, now wherever you install extensions.\n\n\u003e **Simple enough for the vibe-coder novice, powerful enough for the most seasoned dev.**\n\n\u003c!-- savings:hero-hook:start (generated by scripts/build/savings-model.mjs — DO NOT EDIT BY HAND) --\u003e\n### ⚡ Claude Code feeling slow lately? You're not imagining it — and you don't have to live with it. Pool the Claude accounts you already have for **~N× the rate‑limit windows** — 5 Pro accounts match one Max‑20×'s session ceiling for **$115+/mo** back.\n\u003c!-- savings:hero-hook:end --\u003e\n\n[\u003cimg src=\"https://raw.githubusercontent.com/neural-llm-io/power-claude/main/media/marketplace/hero-pays.png\" alt=\"Power Claude pays you back — most Max users never touch their full ceiling. Right-size to the Claude Pro accounts your real usage needs: three pooled Pro accounts billed annually run about $50/mo (plus ~$10/mo for Power Claude) versus $200/mo for Claude Max 20x — about $140/mo back. Pro is ~17% cheaper billed annually, and rotation means no stalls. Start free, $0 for 14 days.\" width=\"100%\"\u003e](https://neural-llm.com/pricing)\n\n\u003c!-- savings:hero-para:start (generated by scripts/build/savings-model.mjs — DO NOT EDIT BY HAND) --\u003e\n\u003e **Save $115+/mo on Claude Code — 5 pooled Pro accounts equal one Max‑20×'s 5‑hour session token ceiling (~220k) for ~58% less** ($83/mo on annual billing vs $200). Claude Code didn't get slower — your single account did. Heavy sub‑agent fan‑out hammers one account's per‑minute limits, every 429 is dead time, and you wait. Power Claude pools the Claude Pro accounts you already pay for and routes every request to the account with the most headroom — so you get **~N× the throughput** (each pooled account is its own independent rate‑limit window) and a 429 never stalls the session (it flips to the healthiest account instantly, no waiting on `Retry‑After`). A watchdog revives crashed, stalled, and rate‑limited sessions in the same tab — while **Token Saver** trims wasteful, repetitive tool‑output token burn and **Token Tree** surfaces where your tokens leak, with your real savings shown live. Cheaper, faster, and it runs through the wall — here's exactly how. ↓\n\u003c!-- savings:hero-para:end --\u003e\n\n```bash\ncode --install-extension neural-llm.power-claude\n```\n\n\u003c!-- install:start (generated from data/install/channels.json by scripts/build/install-channels.mjs — DO NOT EDIT BY HAND) --\u003e\n[![VS Code Marketplace](https://img.shields.io/badge/VS_Code_Marketplace-install-f26522.svg?style=flat-square\u0026logo=visualstudiocode\u0026logoColor=white)](https://marketplace.visualstudio.com/items?itemName=neural-llm.power-claude)\n[![Open VSX](https://img.shields.io/badge/Open_VSX-Cursor_+_VSCodium-a60ee5.svg?style=flat-square\u0026logo=vscodium\u0026logoColor=white)](https://open-vsx.org/extension/neural-llm/power-claude)\n[![npm (CLI)](https://img.shields.io/badge/npm_(CLI)-global_install-cb3837.svg?style=flat-square\u0026logo=npm\u0026logoColor=white)](https://www.npmjs.com/package/power-claude)\n[![Direct .vsix](https://img.shields.io/badge/Direct_.vsix-neural--llm.com-7c3aed.svg?style=flat-square)](https://neural-llm.com/power-claude/downloads)\n[![neural-llm.com](https://img.shields.io/badge/neural--llm.com-downloads_+_docs-7c3aed.svg?style=flat-square)](https://neural-llm.com/power-claude)\n\n**Every way to get Power Claude** — same extension, pick whatever fits your editor and workflow:\n\n| Channel | Best for | Get it |\n|---|---|---|\n| **VS Code Marketplace** | VS Code — one click | `code --install-extension neural-llm.power-claude` |\n| **Open VSX** | Cursor · VSCodium · Gitpod · Windsurf | `cursor --install-extension neural-llm.power-claude` |\n| **npm (CLI)** | any terminal · CI | `npm install -g power-claude` |\n| **Direct .vsix** | air-gapped · manual · pinned version | `code --install-extension power-claude-\u003cversion\u003e.vsix` |\n| **neural-llm.com** | downloads · docs · account | [neural-llm.com](https://neural-llm.com/power-claude) |\n\u003c!-- install:end --\u003e\n\n[![Changelog](https://img.shields.io/badge/changelog-release_notes-6b46c1.svg)](https://neural-llm.com/changelog)\n\n\u003cimg src=\"https://raw.githubusercontent.com/neural-llm-io/power-claude/main/media/marketplace/hero-cost-multiplier.png\" alt=\"Power Claude — N accounts, N× capacity, a fraction of the cost\" width=\"100%\"\u003e\n\n\u003cimg src=\"https://raw.githubusercontent.com/neural-llm-io/power-claude/main/media/marketplace/hero-value-props.gif\" alt=\"Power Claude in action — pool the Claude accounts you already have for N× capacity, zero-stall rate-limit rotation, and real cost savings.\" width=\"100%\"\u003e\n\n\u003e Power Claude is an independent extension by Neural-LLM. It is **not affiliated with Anthropic**; \"Claude\" is a trademark of Anthropic, PBC. Power Claude interoperates with Claude Code's local files — it is third-party tooling.\n\n---\n\n## Stop overpaying for Claude capacity\n\nYou're hitting the wall every afternoon. The obvious fix — upgrade to Max, or start paying per-token API rates — costs a multiple of what you pay now. Power Claude gives you a cheaper path to the same headroom.\n\n**Pool the accounts you already have.** Each Claude.ai Pro or Max subscription is a completely independent organization on Anthropic's side, each with its **own** 5-hour, 7-day, per-minute-token (TPM), and per-minute-request (RPM) rate-limit windows. There is no shared pool between them. Power Claude uses them all, all the time — so **N independent accounts give you roughly N× the capacity on every one of those windows**, at the price of N Pro subscriptions instead of one Max plan or a metered API bill.\n\n\u003c!-- savings:money-block:start (generated by scripts/build/savings-model.mjs — DO NOT EDIT BY HAND) --\u003e\n- 📅 **Annual billing is ~17% cheaper.** Claude Pro is $20/mo month‑to‑month, but $200/yr billed annually — an effective **$16.67/mo, ~17% less**. Annual Pro is the cheapest seat to pool.\n- 💰 **Apples‑to‑apples: 5 Pro = one Max‑20×'s session ceiling, for ~58% less.** The wall you hit mid‑task is the 5‑hour session token ceiling. One Claude Max‑20× tops out around ~220k tokens per session; **5 pooled Pro accounts (~44k each) reach the same ~220k** — for a fraction of the price, and with 5 independent windows instead of 1:\n\n|  | 1 × Claude Max‑20× | 5 × pooled Pro (annual) |\n|---|---|---|\n| 5‑hour session token ceiling | ~220k | ~220k (5 × ~44k) — **same** |\n| Independent rate‑limit windows | 1 | **5** |\n| Monthly cost | $200 | **$83** → **save $117/mo (58%)** |\n\n  *Honest basis:* Anthropic publishes plan multipliers (Max‑20× = \"20× Pro\") and prices, not exact token caps — the ~44k/~220k figures are consistent 2026 estimates. That **20×** is Max‑20×'s *weekly sustained* volume; the **5×** here is the *per‑session ceiling* — the mid‑task wall. Power Claude's own subscription is **$12/mo**, shown separately, never folded into the figures above. Your dashboard runs this against *your* real pool, so it's verifiable — not a slogan.\n\u003c!-- savings:money-block:end --\u003e\n- 📊 **See the savings locally.** Per-account, per-model $/token tracking shows your effective cost under your current subscription tiers versus what the same work would have cost at pay-per-token API rates — so the savings aren't a marketing claim, they're a number in your dashboard.\n- 🎯 **Know when an upgrade is actually worth it.** The cap-hit recommender flags exactly when upgrading a plan would eliminate the throttling that regularly stalls your workflow — and stays quiet when it wouldn't. No upsell guesswork.\n\n\u003cimg src=\"https://raw.githubusercontent.com/neural-llm-io/power-claude/main/media/marketplace/savings-comparison.png\" alt=\"Three ways to buy more Claude capacity — pooled Pro seats vs one Max plan vs per-token API, compared on relative cost and capacity.\" width=\"100%\"\u003e\n\n\u003cimg src=\"https://raw.githubusercontent.com/neural-llm-io/power-claude/main/media/marketplace/demo-cost-analytics.gif\" alt=\"Power Claude cost analytics — per-account, per-model $/token tracking with your effective subscription cost compared against pay-per-token API rates, live in the dashboard.\" width=\"100%\"\u003e\n\n\u003e **Why it works**: Claude.ai Pro and Max subscriptions each carry their own 5-hour, 7-day, TPM, and RPM rate-limit windows. Accounts are completely independent organizations on Anthropic's side — there is no shared pool. Power Claude simply uses them all, all the time. It does **not** bypass, circumvent, or violate any Anthropic rate limit or term of service — every account stays within its own published limits.\n\nAll cost math runs locally. Power Claude reads Claude Code's own local state and session transcripts — the same files Claude Code writes. No token counts, transcripts, or usage data leave your machine.\n\n---\n\n## Power Claude vs. vanilla Claude Code\n\n| | Claude Code alone | With Power Claude |\n|---|---|---|\n| **Rate limit (429)** | Wait for the 5-hour / weekly window to reset | Auto-rotates to the account with the most headroom — N accounts ≈ N× throughput |\n| **Cost of more capacity** | Upgrade to Max, or pay per-token API rates | Pool only the Pro accounts your usage needs — *N* accounts ≈ *N*× a single account, in *N* independent windows |\n| **Cost visibility** | None built in | Per-account, per-model $/token tracking + subscription-vs-API comparison |\n| **Multiple accounts** | Manual logout / login each time | One pool, automatic routing (Balanced Mode scores every account per request) |\n| **Concurrent sessions** | Serialized against one account | Spread across the pool — a round finishes ~N× faster |\n| **Crashed / stalled session** | Notice late, restart by hand, lose context | Watchdog detects crashed/stalled/rate-limited sessions + one-click auto-resume |\n| **Session history** | Hunt through JSONL transcripts + git diffs | Insight Graph: files touched, model, cost, completion status per session |\n| **AI-edit blame** | None | Editor-gutter blame per Claude session (runs alongside GitLens) |\n| **Where your data goes** | — | All local — no prompts or code ever leave your machine |\n| **Editors** | VS Code, Cursor | VS Code, Cursor, and a CLI — one shared `~/.power-claude/` |\n\n---\n\n## Core capabilities\n\n### 🔄 Multi-account rotation + smart 429 routing\n\nConnect personal, team, and project-specific Claude accounts. Power Claude watches rate limits in real time via the Claude Code state directory that Claude Code writes itself. When the active account hits a cap:\n\n1. The Stop hook fires and Power Claude detects the rate-limit signal\n2. The soonest-recovering account is activated automatically\n3. The interrupted session resumes in the same VS Code tab — same context, same conversation, no retyping\n\n**Smart 429 routing** is the proactive half. When an account returns 429 — whether a per-minute throttle or a hard quota signal — the proxy tries a healthy alternate first instead of sleeping up to 120 seconds on `Retry-After`. The throttled account is put in cooldown so the selector can't pick it again until Anthropic says it's ready. **Anti-ping-pong is structural, not a heuristic**: each 429 marks its account out for its `Retry-After` window, so the proxy can't bounce between two throttled accounts. If the entire pool is throttled at once, it falls back to the existing wait-and-retry path on the least-bad account — you never end up worse off than before.\n\nThe status bar shows a live count of healthy / cooling / blocked accounts and the cumulative `N throttle waits avoided`. The accounts dashboard shows cooldown ETAs and lets you manually activate, refresh, or re-authenticate any account.\n\n\u003cimg src=\"https://raw.githubusercontent.com/neural-llm-io/power-claude/main/media/marketplace/demo-rotation-zero-wait.gif\" alt=\"Power Claude rate-limit rotation in action: the active account hits a 429 with a cooldown timer and Power Claude auto-switches the in-flight request to the healthiest account in the pool — work continues with zero wait.\" width=\"100%\"\u003e\n\n\u003cimg src=\"https://raw.githubusercontent.com/neural-llm-io/power-claude/main/media/marketplace/08-balanced-mode-status-bar.png\" alt=\"Power Claude status bar showing the live \"Balanced · 3 accounts · ~3× throughput\" indicator with per-account health dots.\" width=\"100%\"\u003e\n\n### 🗂️ Account groups (isolated context)\n\nPartition Claude accounts into named groups (Work / Personal / clients) with **isolated** projects/memory trees so context does not bleed across roles.\n\n- Create groups in the dashboard (**Profiles → Account groups**) or `pc scope create`\n- Drag accounts between groups, or `pc scope add` / `pc scope move`\n- **Use this group** / `pc scope use` selects the active set for new work\n- Credentials stay private per account — groups organize membership and local context only\n\n### ⚡ Power Mode — N× throughput\n\nMulti-account rotation reacts to exhaustion. Power Mode is proactive: every request is routed to the **least-busy account simultaneously**, so all your accounts fill in parallel instead of draining one before moving to the next. With three accounts you get an effective **~3× ceiling** on each independent 5h / 7d / TPM / RPM window — and the per-minute throttling that used to choke single-account streaming during heavy sub-agent fan-out simply doesn't happen.\n\n| Mode | How it works | Best for |\n|------|-------------|----------|\n| **Standard rotation** | Stays on one account until it exhausts, then switches | Light to moderate usage |\n| **Power Mode** | Every request goes to the least-loaded account | Heavy / continuous coding sessions |\n\nThe status bar shows `⚡ Balanced [3] ~3×` while active, so you can see the multiplier at a glance. Enable from the **Power Claude Settings** panel in one click — no CLI, no config files. Or via CLI: `power-claude balanced on`.\n\n\u003cimg src=\"https://raw.githubusercontent.com/neural-llm-io/power-claude/main/media/marketplace/demo-parallel-throughput.gif\" alt=\"Power Claude Power Mode — every request is routed to the least-busy account simultaneously, filling all accounts in parallel for ~N× throughput during heavy sub-agent fan-out.\" width=\"100%\"\u003e\n\n### ⏱ Runout Forecaster — see the wall coming hours in advance\n\nPower Mode and smart 429 routing prevent most rate-limit hits, but they can't *stretch* your subscriptions past their real limit. The Runout Forecaster watches utilization across every account and projects, at your current burn rate, whether you'll exhaust capacity before the 5-hour or 7-day windows reset.\n\n| State | Banner |\n| ----- | ------ |\n| Safe | **green** — \"Pool safe through next reset (~3h 41m)\" |\n| Some accounts at risk | **amber** — \"3 of 12 accounts at risk\" |\n| Pool exhaustion imminent | **red** — \"Pool exhausts in ~28m\" |\n\nEach account row also gets a tiny per-account ETA chip (`✓ safe`, `ETA 1h 47m`, `ETA 14m`) so you can spot the bottleneck without opening drawers.\n\n**The math is honest**: reset times come from Anthropic's authoritative rate-limit headers, so the \"will I make it before reset?\" half of the calculation is *exact*. Only burn rate is estimated, from a rolling sample buffer of the last 30 minutes of utilization observations. When there isn't yet enough data to project (first ~2 minutes after activation, or no activity), the banner is suppressed and the chips show `—`.\n\n### ↻ Auto-resume + resilience watchdog + Auto-Resume Nudges\n\nRate limits aren't the only thing that kills a session. Process kills, OOM events, IDE crashes, network drops mid-tool-call, and sessions that ended cleanly but left work unfinished all used to leave you with a stale tab and a manual recovery.\n\nThe watchdog polls every 30 seconds and applies four detection rules:\n\n| Trigger | Signal | What it means |\n|---------|--------|---------------|\n| **Heartbeat-stale** | No `lifecycle:heartbeat` in N seconds AND JSONL idle | Process kill, OOM, IDE crash, or network drop |\n| **Tool-truncation** | Last assistant record is `stop_reason: tool_use` with no follow-up | Died mid-tool-call |\n| **Ended with pending tasks** | Stopped cleanly but task items still pending | Session said \"done\" but work remains |\n| **Rate-limit** | `pending-recovery.json` written by Stop hook | Standard rate-limit rotation |\n\nWhen the watchdog fires, Power Claude surfaces a one-click QuickPick with a reason-aware resume prompt — each death type gets tailored wording so the agent picks up with the right context.\n\n**Auto-Resume Nudges** catch the more annoying case: sessions that stop *politely* when they should keep going. Claude pauses to ask \"want me to continue?\", offers a three-option menu when one option is obvious, or tells *you* to run a command it could have run itself. A family of toggleable detectors fires on every Stop event; when a pattern matches and no hard-stop veto fires (auth failure, missing credential, destructive op pending, approval gate), the Stop hook re-issues the turn with a *targeted* corrective prompt:\n\n| Default-on nudge | What it catches | What the agent sees back |\n|------------------|-----------------|--------------------------|\n| `ask-permission-to-continue` | \"Want me to continue?\" / \"Should I keep going?\" | \"Continue the work. Don't ask permission for obvious next steps.\" |\n| `offer-multi-choice` | \"Three reasonable next moves — 1) … 2) … 3) … which?\" | \"Pick the most obvious option and proceed.\" |\n| `tell-user-to-do-work` | \"You can run X to verify\" / \"You'll need to deploy\" | \"You have the tools to do this yourself.\" |\n| `pseudo-checkpoint` | \"Good stopping point\" / \"Let's pause here\" with no reason | \"There's no reason to pause. Continue to the next concrete deliverable.\" |\n\nFour more nudges ship opt-in for aggressive cases. A **shadow mode** logs every would-fire decision without resuming so you can audit before going live, and a **loop guard** caps consecutive nudges at 3. An **optional LLM fallback** (off by default; Haiku 4.5 when on) handles ambiguous stops the deterministic catalog missed — bounded to one ~128-token classification call per Stop, empirically ~3% of stops. The catalog lives at `~/.power-claude/config/nudges.json`; edit it directly to tune patterns.\n\n\u003cimg src=\"https://raw.githubusercontent.com/neural-llm-io/power-claude/main/media/marketplace/demo-watchdog.gif\" alt=\"Power Claude resilience watchdog in action — it detects a session interrupted mid-task and offers one-click recovery (resume in a healthy account, open read-only, or dismiss), so a 429 mid-task never loses the run.\" width=\"100%\"\u003e\n\n### 🩹 Session Actions — one-click recovery, right on the line that broke\n\nWhen a turn dies with **\"Prompt is too long\"**, you don't have to copy context into a fresh chat by hand. Power Claude detects the error on the chat line and injects a branded **Heal** button right there — click it and the session is trimmed and resumed automatically (the same trim-and-resume recovery the watchdog uses), no retyping.\n\nHeal is the first of a family of **Session Actions** — recovery actions matched to *how* a session got stuck:\n\n| Action | When it appears | What it does |\n|--------|-----------------|--------------|\n| **Heal** | \"Prompt is too long\" / compaction thrash | Trim the conversation and resume in a fresh session |\n| **Unstick** | Repeating tool loop | Fork from the last good checkpoint |\n| **Revive** | Crashed / abandoned | Fork from the last good checkpoint |\n| **Re-run tool** | Truncated mid tool-call | Re-issue the tool call |\n| **Finish tasks** | Stopped with TODOs pending | Resume and inject the pending tasks |\n| **Resume** | Idle / stuck after a rate-limit window | Plain resume |\n| **Repair turn** | Wedged thinking turn (HTTP 400 on resume) | Trim the orphaned turn |\n\nEvery action is available three ways: the **inline button** on the chat line, the **`Power Claude: Session Actions`** command (pick a session → pick an action), and the **CLI** (`pc session action heal \u003csession-id\u003e`). Each action is settings-gated (`powerClaude.sessionActions.*`) and its recovery strategy is configurable. Manual by default — you stay in control.\n\n### 📊 Insight Graph + cost analytics + editor-gutter blame\n\nAfter a long session, you want to know: *What did it change? Which files? What model, and what did it cost? Did it finish?* Today you answer those by hunting through JSONL transcripts and git diffs. Power Claude answers them in one click, rendering your session history as a commit-graph-style lane view — the visual language GitLens made familiar for git, applied to AI coding sessions.\n\n\u003cimg src=\"https://raw.githubusercontent.com/neural-llm-io/power-claude/main/media/marketplace/hero-insight-graph.png\" alt=\"Power Claude Insight Graph\" width=\"100%\"\u003e\n\nEach row is a session, carrying a model badge (with version), total cost, files-changed count, the account that handled it, and a completion status (done / pending / untracked). Click any session to expand the full detail without leaving VS Code:\n\n- **Files panel** — every file read, written, or edited, with per-file diff vs the session's git snapshot, op-type counts, and timestamps\n- **Sub-agents** — forked agents appear as child lanes linked to the parent\n- **Hooks fired** — which Claude Code hook scripts ran, with exit codes and timing\n- **Rate-limit events** — when the session hit a cap, which account it rotated to, and the cooldown length\n- **Plans drafted** and **commits authored** — plan-mode artifacts and every `git commit`, clickable to open the diff\n- **Multi-track timeline** — tool calls, hook firings, model switches, and rate-limit events on one synchronized time axis\n\n**Cost analytics** track token consumption and dollar cost per session, per account, and per model with plan-tier-aware math: per-account breakdowns by model with daily / weekly / monthly rollups, anchor markers showing exactly when caps were hit, the cap-hit upgrade recommender, and your subscription savings versus pay-per-token API rates.\n\n**Inline session blame** annotates each region of any open file with the Claude session that last touched it — exactly like git blame, but for AI editing activity. Hover for the session summary (date, model, cost, completion status); click to jump to the full detail view. It runs alongside GitLens in a separate gutter column with distinct visual language, so the two never conflict.\n\n\u003cimg src=\"https://raw.githubusercontent.com/neural-llm-io/power-claude/main/media/marketplace/10-session-blame-gutter.png\" alt=\"Power Claude session-blame gutter: each block of code is annotated with the Claude session that authored it, its model, and cost, with a hover card showing full session detail.\" width=\"100%\"\u003e\n\n### 🎯 Preserve-Focus Fix — stop Claude Code from yanking your cursor\n\nYou're typing in a file. Claude finishes a response. Suddenly your cursor jumps to the Claude panel and your keystrokes land in the wrong place — or you accidentally accept a pending action because your finger was already on `Enter`.\n\nThis is [anthropics/claude-code#32726](https://github.com/anthropics/claude-code/issues/32726) — open since early 2026, no shipping fix. The Claude Code VS Code extension calls `WebviewPanel.reveal()` without `preserveFocus: true` on six code paths. Power Claude ships the patch Anthropic hasn't:\n\n- **Two settings** — `powerClaude.preserveFocusFix.enabled` (default OFF — opt in once) and `powerClaude.preserveFocusFix.autoRunOnUpgrade` (default ON). Flip the first to `true`, reload the window, and the focus-steal is gone.\n- **Six surgical patches** to the installed `extension.js`, anchored on stable string literals; after every Claude Code upgrade Power Claude detects the changed mtime and re-patches automatically.\n- **Idempotent and reversible** — a sentinel comment makes re-runs a no-op, and a one-shot `.preserve-focus.orig` backup beside each patched file lets you undo cleanly.\n- **Manual command** — `Power Claude: Re-apply Preserve-Focus Patch to Claude Code Extension` for the rare case the auto-detector misses an upgrade.\n\nIf Anthropic ships the proper upstream fix, you'll uninstall this in one toggle.\n\n---\n\n### 🎨 Purple Scrollbars — Power Claude's brand, on your scrollbars\n\nA small one: Power Claude themes VS Code's scrollbars in its own brand violet, on by default.\n\n- **One setting** — `powerClaude.scrollbarTheme.enabled` (default ON). VS Code has no per-extension or per-tab scrollbar API, so this necessarily colors scrollbars across the whole editor, not just Power Claude's own views — turn it off any time via Settings → Power Claude: 🎨 Branding \u0026 Theme.\n- **Non-destructive** — your prior `workbench.colorCustomizations` values (or the absence of any) are snapshotted before the first apply and restored exactly on disable or uninstall.\n- **No reload needed** — applies and reverts live.\n\n---\n\n### ⬆️⬇️ Chat scroll buttons — jump to the top or bottom of a long conversation\n\nSmall purple floating **↑ / ↓** buttons on the right edge of the *Claude Code\nchat thread*: one click jumps straight to the beginning of the conversation,\none click jumps to the newest message — no hand-scrolling through a\nthousand-line thread.\n\n- **One setting** — `powerClaude.chatScrollNav.enabled` (default ON, Settings →\n  Power Claude: 🧩 Claude Core).\n- **Only when useful** — the buttons appear only while the thread is actually\n  long enough to scroll; short chats stay clean.\n- **Smooth scrolling** that respects your OS's reduced-motion preference, and\n  survives Claude Code auto-updates via Power Claude's self-heal.\n\n---\n\n### 🧪 Detached Sessions *(experimental · Pro)* — your turn keeps running when you close the tab\n\nLove that the Claude **CLI** keeps working when you refresh or close the window — but want a real **graphical UI** instead of a bare terminal? That's Detached Sessions.\n\nTurn it on and a turn you launch runs inside its own **systemd user service** — its own cgroup, the same escape the rotation proxy relies on — so a browser-tab sleep, a page refresh, or closing the window **can't kill it**. When you come back, the dashboard **reattaches and replays** everything you missed, then live-tails the rest. The CLI's persistence, behind the GUI.\n\n- **Resilience** — stop losing a 10-minute agent run to a flaky code-server / remote connection. It finishes whether or not you're watching.\n- **Efficiency** — reconnect to the *same running work* instead of burning a fresh turn to rebuild context; you don't pay twice for the same thing.\n- **Safe by construction** — default OFF; every turn still routes through the rotation proxy (quota \u0026 savings tracking intact); bounded by a hard turn-timeout, an idle reaper, and a concurrency cap; honours the `pc emergency` kill-switch.\n- **Requires** Linux with a `systemd --user` manager (typical for code-server / remote dev). On other hosts the toggle stays disabled — it never pretends to persist when it can't.\n\nEnable it from the dashboard's **Rotation Safety Controls** or the **Setup \u0026 Optimize** wizard.\n\n---\n\n## Quickstart\n\nTwo install paths, both fully working from a single command.\n\n**VS Code Marketplace (recommended):**\n\n```bash\ncode --install-extension neural-llm.power-claude\n```\n\nOr search \"Power Claude\" in the VS Code Extensions panel. After install, click the Power Claude icon in the activity bar. The extension creates `~/.power-claude/` automatically, migrates any legacy data, and starts the local rotation proxy. No PATH changes required — every action is available from the dashboard and command palette.\n\n**npm CLI (terminal-only or scripting):**\n\n```bash\nnpm install -g power-claude\npower-claude install\n```\n\n`npm install -g` puts two binaries on your PATH: `power-claude` (canonical) and `pc` (short alias). `power-claude install` creates `~/.power-claude/`, deploys lib + hook scripts, and appends the bin path to your shell rc. Use `pc help` to see the full command list.\n\n**Commands at a glance:**\n\n```bash\npc list                       # all accounts with quota bars + status\npc status                     # proxy status dashboard\npc watch                      # LIVE TUI — account table + activity stream + modal hotkeys\npc rotate                     # rotate to next healthy account\npc rotate work-account        # rotate to specific account\npc health                     # probe accounts and render with live data\npc add                        # add an account profile\npc switch \u003cname\u003e              # switch active account\npc relogin \u003cname\u003e             # re-authenticate one account (OAuth flow)\npc save-profile \u003cname\u003e        # save current credentials as named profile\npc remove \u003cname\u003e              # delete a profile\npc export --output backup.json # export/back up all accounts to a file\npc import backup.json         # restore / move accounts onto another machine\npc logs -f --level error      # tail event log with filters\npc run                        # launch `claude` with auto-rotation\npc serve --port 8888          # start proxy server\npc doctor                     # full diagnostic checklist\npc config get|set|list        # CLI configuration\npc disable [--strip-profiles] # turn off rotation (destructive variant strips non-active)\npc enable                     # turn rotation back on\n```\n\n**Both install paths use the same runtime directory** (`~/.power-claude/`), so you can have the extension AND the CLI installed and they share state.\n\n**Add more accounts:** click **Link Account** in the accounts tree (extension) or run `pc relogin \u003cname\u003e` (CLI) to walk through `claude login` for a new account profile. Multiple accounts enable rotation. See [Pooling more accounts — one inbox, many logins](#pooling-more-accounts--one-inbox-many-logins) for the fastest way to stand several up.\n\n**Move accounts to another workspace or computer:** set everything up once, then carry the whole pool over instead of re-logging in everywhere. In the extension, use the **Export** / **Import** buttons on the Accounts panel; from a terminal, `pc export --output accounts.json` then `pc import accounts.json` on the other machine. The bundle contains your login credentials in plain text — keep it private, move it over a trusted channel, and delete it once imported (or use `pc export --gpg \u003ckey\u003e` for an encrypted bundle). Accounts that already exist on the destination are skipped automatically.\n\n### Pooling more accounts — one inbox, many logins\n\nEach Claude account is a separate Anthropic subscription with its **own** rate-limit windows — that independence is exactly what Power Claude pools. Every account you add is a real, paid subscription (Pool ≠ free capacity); the trick below is just how to manage many of them without juggling many inboxes. Every account stays within its **own** published limits — nothing here bypasses or circumvents Anthropic's terms.\n\nYou don't need a separate inbox per account. Sign each one up with its own **email alias** that all forward to a single inbox, then `claude login` each as a Power Claude profile:\n\n| How | What you do | Cost |\n|---|---|---|\n| **Plus-addressing** | `you+pc1@gmail.com`, `you+pc2@gmail.com`, … all land in `you@gmail.com` (Gmail, Fastmail, iCloud, most providers). | Free, instant — but some signup forms reject the `+`. |\n| **Custom-domain aliases / catch-all** | Own a domain? Route `pc1@yourdomain.com`, `pc2@yourdomain.com` (or a `*@` catch-all) to your inbox. [Cloudflare Email Routing](https://developers.cloudflare.com/email-routing/) is free; Google Workspace and Fastmail support it natively. | Free–cheap; most reliable for signups. |\n| **Alias service** | [SimpleLogin](https://simplelogin.io/), [addy.io](https://addy.io/), Apple **Hide My Email**, or Firefox Relay mint unlimited forwarding aliases on demand. | Free tiers cover plenty. |\n| **Separate real inboxes** | Just use distinct real email addresses. | Free. |\n\nIf you run your own mail server, a single admin command per alias is enough — e.g.:\n\n```bash\n# one alias per account, all forwarding to your real inbox\nsetup alias add power-claude-1@yourdomain.com you@yourdomain.com\nsetup alias add power-claude-2@yourdomain.com you@yourdomain.com\n# …repeat, then `claude login` each at https://claude.ai and save it as a profile\n```\n\nThen add each in Power Claude (**Link Account**, or `pc relogin \u003cname\u003e`). Pro accounts on **annual billing** are the cheapest seats to pool — see [Stop overpaying](#stop-overpaying-for-claude-capacity).\n\n**Enable the full resilience watchdog (optional, extension only):** run \"Power Claude: Install Session Lifecycle Hooks\" from the command palette. Three lightweight hooks land in `~/.power-claude/hooks/lifecycle/` and emit heartbeat / session-end markers; the watchdog's precision detection depends on these. Opt-in; nothing in your Claude Code config is modified without explicit consent.\n\n**Enable editor gutter blame (optional):** run \"Power Claude: Install Timeline Hook\" from the command palette to enable AI-edit annotations in VS Code's Timeline view and the editor gutter.\n\n**Pro features:** enter your license key in `Settings \u003e Power Claude \u003e License Key`. A 7-day free trial starts automatically on install — no card required. See [Free Trial vs Pro](#free-trial-vs-pro) for what each tier includes.\n\n---\n\n## Free Trial vs Pro\n\nSee [neural-llm.com/pricing](https://neural-llm.com/pricing) for current pricing and tier details.\n\n**Free to download and install.** On first run Power Claude automatically starts a 7-day free trial — no card required, full Pro access for the window. After the trial, paid features (proxy, multi-account rotation, parallel dispatch, account pinning, the watchdog) lock and the extension runs in passthrough mode; only the read-only dashboard, profile management, and session resume remain. There is no ongoing free tier.\n\nPro unlocks unlimited accounts, the full watchdog with all four detection rules, the editor gutter blame annotations, the session graph view (Cytoscape.js force-directed and hierarchical layouts), pre-commit review, advanced search with field filters, and session history with no day limit.\n\n**Try Pro with founder pricing — the best value.** Start a **14-day Premium Trial** at [neural-llm.com/pricing](https://neural-llm.com/pricing): card on file, **$0 today**, and your first charge isn't until day 15 — at the discounted **founder rate you lock in for 2 full years**. Cancel anytime before day 15 and you're never charged. Prefer no card at all? The 7-day free trial (no card) starts automatically on install.\n\n[\u003cimg src=\"https://raw.githubusercontent.com/neural-llm-io/power-claude/main/media/marketplace/premium-trial-cta.png\" alt=\"Start your 14-day Premium Trial — $0 today, cancel anytime before day 15\" width=\"100%\"\u003e](https://neural-llm.com/pricing)\n\n---\n\n## FAQ\n\n**Does this work with the Anthropic API directly, or only with Claude Code?**\n\nPower Claude is built specifically for Claude Code — the CLI and IDE tool at `claude.ai/code`. It reads the local state Claude Code writes under its own config directory. It does not work with raw Anthropic API calls or with Claude.ai in the browser.\n\n**What data leaves my machine?**\n\nOnly your license key, a hashed machine identifier, and the extension version are transmitted — to `api.neural-llm.com` during license validation. That validation is cached locally for 24 hours; no call is made on every keystroke. No prompts, no file contents, no transcripts, no usage data, no authentication tokens. Full details at [neural-llm.com/privacy-policy](https://neural-llm.com/privacy-policy).\n\n**Can I use multiple Claude accounts?**\n\nYes. This is one of Power Claude's core capabilities. Add as many accounts as you have. Power Claude tracks health, cooldown state, and spend per account, and rotates automatically when the active account hits a rate limit. The Insight Graph annotates which account handled each part of a session.\n\n**Can I run several accounts from one email inbox?**\n\nYes — you don't need a separate inbox per account. Sign each Claude account up with its own email *alias* (plus-addressing like `you+pc1@gmail.com`, a custom-domain catch-all, or an alias service such as SimpleLogin / iCloud Hide My Email), all forwarding to one inbox, then add each as a Power Claude profile. Each account is still its own real, paid Anthropic subscription within its own limits — the alias is purely for managing many logins from one place. Step-by-step: [Pooling more accounts — one inbox, many logins](#pooling-more-accounts--one-inbox-many-logins).\n\n**What happens when I hit a Claude rate limit?**\n\nIf you have a second healthy account, Power Claude rotates to it automatically and resumes your session — usually within a few seconds. If all accounts are in cooldown, Power Claude displays the soonest-recovering countdown and optionally (`powerClaude.autoSelectSoonest`) activates the first account whose cooldown clears.\n\n**How is this different from GitLens?**\n\nGitLens maps your git history to your code. Power Claude maps your *AI session history* to your code. The two are complementary and designed to coexist. GitLens tells you which commit introduced a line. Power Claude tells you which Claude session wrote that commit, what it cost, which model it used, and whether it finished what it started.\n\n**How much does Pro cost?**\n\nSee [neural-llm.com/pricing](https://neural-llm.com/pricing) for current pricing. A 7-day free trial (no card) starts automatically on install. The 14-day Premium Trial (card on file, $0 today, first charge on day 15, cancel anytime before then) locks in founder pricing for 2 years. There is no ongoing free tier — after the trial a paid subscription is required for Pro features.\n\n**What is the refund policy?**\n\nFull refund within 14 days of purchase, no questions asked. Contact [support@neural-llm.com](mailto:support@neural-llm.com) with your order details.\n\n---\n\n## Extending Power Claude\n\nPower Claude exposes a stable public API for other VS Code extensions to read session data, contribute custom tabs to the Session Detail webview, and enrich the Insight Graph with domain-specific data.\n\n```ts\n// In your extension's activate():\nconst powerClaude = vscode.extensions.getExtension(\"neural-llm.power-claude\");\nconst api = await powerClaude.activate();   // PowerClaudePublicApi\n\n// Read the live session index.\nconst sessions = api.sessions.list({ limit: 20 });\napi.sessions.onDidChange(() =\u003e /* re-read */);\n\n// Contribute a new tab to the Session Detail webview.\napi.contributeSessionDetailTab({\n  id:    \"myExt.audit\",\n  label: \"Audit\",\n  render(ctx) {\n    return {\n      html: `\u003cp\u003eAudit findings for ${ctx.sessionId}\u003c/p\u003e`,\n      messageHandlers: {\n        \"myExt.audit.openTrace\": (payload) =\u003e /* … */,\n      },\n    };\n  },\n});\n\n// Contribute a session enricher (lightweight read-only augmentation).\napi.contributeSessionEnricher({\n  id: \"myExt.severityScore\",\n  label: \"Severity Score\",\n  enrich(session) { return { score: 0.42 }; },\n});\n```\n\nFull API surface at [neural-llm.com/power-claude](https://neural-llm.com/power-claude). Contracts under the `PowerClaudePublicApi` interface follow semver — breaking changes are rare and announced in the changelog.\n\n---\n\n## What Power Claude does NOT do\n\n- Does not make API calls to Anthropic on your behalf\n- Does not bypass, circumvent, or violate any Anthropic rate limit or term of service — every account stays within its own published limits\n- Does not transmit your prompts, code, transcripts, or authentication tokens off your machine\n- Does not modify your global Claude Code config without explicit consent (all hook installers are opt-in command-palette invocations)\n- Does not require any VS Code workspace to be open — the extension activates on Claude Code state changes regardless\n\n---\n\n## What Power Claude writes to disk\n\nEverything runtime lives under **`~/.power-claude/`**:\n\n| Path | Contents |\n|------|----------|\n| `~/.power-claude/profiles/` | OAuth profiles for each Claude account (rotation pool) |\n| `~/.power-claude/state/` | Rotation state, lockouts, pending-recovery markers |\n| `~/.power-claude/logs/` | events.jsonl, watchdog.log, rotation-audit.log |\n| `~/.power-claude/lib/` | Bundled shell helpers, deployed at install time |\n| `~/.power-claude/hooks/` | Lifecycle + rate-limit hook scripts |\n| `~/.power-claude/cache/` | devtools-data.json, session-index.json |\n| `~/.power-claude/config/` | config.json, auto-rewake-config.json |\n\nThe **only** files Power Claude writes outside `~/.power-claude/`:\n\n| File | Why |\n|------|-----|\n| Claude Code's credentials file | Claude Code reads its OAuth credentials from here. Power Claude overwrites this file on rotation so Claude Code sees whichever account is active. |\n| Claude Code's settings file | Hook registrations (with your consent). Power Claude appends entries; entries from other tools and your own settings are preserved. Removing the extension via \"Power Claude: Uninstall Everything\" strips every Power Claude entry cleanly. |\n\nEarlier versions stored profiles, state, and hook scripts under Claude Code's config directory. On first run, the migration module moves any legacy data into `~/.power-claude/` and deletes the legacy copies — no manual cleanup needed.\n\n**Complete uninstall:** run \"Power Claude: Uninstall Everything\" from the command palette, or `rm -rf ~/.power-claude \u0026\u0026 code --uninstall-extension neural-llm.power-claude`. The uninstall command strips every Power Claude entry from Claude Code's settings file and removes `~/.power-claude/`; your Claude Code credentials and other settings are untouched.\n\n---\n\n## Privacy\n\nLocal-first. Power Claude reads from:\n\n- `~/.power-claude/profiles/` — your rotation pool's OAuth profiles\n- `~/.power-claude/state/` — rotation state, lockouts, recovery markers\n- `~/.power-claude/logs/` — events.jsonl, rotation-audit.log\n- Claude Code's own session transcripts — used only for token-bucket counts and session indexing (read-only)\n- Claude Code's edit-history state — only if you opt in to Timeline tracking\n\nThe **only** files Power Claude writes outside `~/.power-claude/` are Claude Code's credentials file (so Claude Code picks up the rotated account) and Claude Code's settings file for hook registrations (with your consent). See **What Power Claude writes to disk** above for the complete list.\n\nNo data leaves your machine. No cloud telemetry. Full privacy policy at [neural-llm.com/privacy-policy](https://neural-llm.com/privacy-policy).\n\n---\n\n## Get Power Claude\n\n```bash\ncode --install-extension neural-llm.power-claude\n```\n\n**Two ways to start, both $0 today:** the **7-day free trial** (no card, auto-starts on install), or the **14-day Premium Trial** that locks in founder pricing for 2 years (card on file, nothing charged until day 15, cancel anytime). Start pooling your accounts today and watch the savings add up in your dashboard.\n\n[![Start your 14-day Premium Trial — $0 today, cancel anytime before day 15](https://img.shields.io/badge/Start_your_14--day_Premium_Trial-%240_today-22c55e?style=for-the-badge\u0026labelColor=16a34a)](https://neural-llm.com/pricing)\n\n---\n\n## Support\n\n- Bug reports: [github.com/neural-llm-io/power-claude/issues](https://github.com/neural-llm-io/power-claude/issues)\n- Email: [support@neural-llm.com](mailto:support@neural-llm.com)\n- Sales / licensing: [support@neural-llm.com](mailto:support@neural-llm.com)\n\n---\n\n## License\n\nNeural-LLM commercial EULA — see [LICENSE](LICENSE) and [neural-llm.com/eula](https://neural-llm.com/eula). Copyright (c) 2025-2026 Neural-LLM.\n\n---\n\n## Trademarks\n\nClaude and Claude Code are trademarks of Anthropic PBC. Visual Studio Code is a trademark of Microsoft Corporation. Cursor is a trademark of Anysphere Inc. Neural-LLM is not affiliated with, endorsed by, or sponsored by Anthropic, Microsoft, or Anysphere. Power Claude is an independent third-party tool that interoperates with the public file outputs of Claude Code.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fneural-llm-io%2Fpower-claude","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fneural-llm-io%2Fpower-claude","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fneural-llm-io%2Fpower-claude/lists"}