Power Claude has shipped a lot in the past few months. The version-by-version changelog tells the literal story; this is the narrative version — what each release was trying to solve, why we shipped it in that order, and which improvements have produced the most real-world impact for developers running Claude Code under heavy load.
The proxy core — version 0.15.0
The single release that mattered most was 0.15.0. Before it, rotation was a watchdog-driven swap of the active credential after a 429 had already landed. The user saw the error briefly, the rotator triggered, and the next prompt worked. The mechanism was correct but the user experience was visibly interrupted.
0.15.0 replaced that model entirely. The new proxy is a small Node.js process that sits between Claude Code and api.anthropic.com. Every API call passes through it. The proxy reads Anthropic's rate-limit headers on every response — anthropic-ratelimit-unified-5h-utilization and anthropic-ratelimit-unified-7d-utilization — and rotates accounts pre-emptively when utilization crosses a configurable threshold (default 98%). The 429 that the old model surfaced briefly is now structurally avoided. The client never sees the failure.
The architectural shift had cascading effects:
- SSE streaming passthrough. The proxy pipes streamed responses unbuffered. There is no detectable latency overhead on streaming completions.
- 429 transparent retry. Even when a 429 does land — for cases the pre-emptive logic missed — the proxy retries on the next account immediately. The client still does not see it.
- OAuth token auto-refresh. The proxy refreshes access tokens five minutes before expiry, with concurrent-refresh coalescing so parallel requests do not trigger multiple refreshes.
What 0.15.0 produced for users is the "rotation that you never see" experience. It is the headline behavior that made the rest of the extension's surfaces useful — once 429s stopped interrupting, the session resilience and cost analytics had room to matter.
Parallel account mode
The proxy also unlocked parallel account distribution. Before 0.15.0, the rotator served all of one session's traffic from one account at a time. Even with a pool of five accounts, only one was ever actively serving requests; the others were warm spares.
Parallel mode changes the dispatch model: requests are distributed across the pool simultaneously. With five MAX-tier accounts in parallel, effective Opus throughput is roughly ten million tokens per minute (against approximately two million from one account). Two dispatch strategies ship:
round-robin— sequential, predictable. Each request goes to the next account in the pool. Useful for simple workloads where parallelism is the only goal.least-utilized— routes each request to the account with the lowest current 5h utilization plus in-flight request weight. The default. Produces the most even pool utilization under sustained load.
Parallel mode is opt-in (powerClaude.proxy.parallelAccounts: true). The configurator lets users toggle accounts in and out of the pool, switch strategies, and preview the estimated throughput multiplier. Hot-reload means no proxy restart.
The user-visible effect is throughput that scales with pool size, not with the active-account ceiling. Fan-outs that used to trip RPM on a single account now distribute across the pool naturally. Parallel dispatch ships in the extension itself, free to download and try, no credit card needed.
Account pinning and tagging
A late-stage feature in the 0.15.0 release was pinning and tagging — the controls for users who wanted to influence which accounts the rotator preferred without writing custom router logic.
pinnedAccountsrestricts rotation to a named subset of accounts. Use case: a developer with three Pro accounts and two Max accounts pinning to the Max set so Opus traffic stays on the higher-headroom tier.pinnedTiersdoes the same selection by tier rather than by account name. Same effect, less brittle when accounts come and go.- Account tags let users label accounts (
work,personal,experiment) and gate features by tag. Particularly useful when a developer wants to keep an experiment account out of the main pool to track its usage separately.
The combination produced a configuration surface that handles most "I want to favor account X" workflows without code changes.
Confirm-gated auto-resume — version 0.16.0
0.16.0 was the auto-resume release. The watchdog had been capturing session state since the early versions, but the "resume from captured state" UX had not graduated past manual click. 0.16.0 added two modes that triggered automatically.
The first was the basic lazy-stop catcher. Many Claude Code sessions end with the model saying something like "want me to continue?" and stopping. The watchdog now recognizes that pattern on a Stop hook and re-issues an affirmative continuation prompt automatically. No human re-prompting required.
The second mode was the more interesting one: the confirm-gated complete-remaining-tasks nudge. When the model finishes one piece of work and lists remaining tasks that were part of the original ask, the watchdog detects the pattern but cannot safely resume blindly — the "remaining tasks" list might hide a real decision point. So the nudge routes the matched message and the session's original request through a Haiku call to verify the remaining work is a trivial continuation. Only when Haiku confirms does the watchdog resume the session.
Two other 0.16.0 features that shipped alongside the confirm-gated nudge:
shadowOnDisablediscovery telemetry. A disabled nudge can still record "would have fired" events. The extension learns which nudges would have helped the user, even when off.- Feature-discovery onboarding. After enough "would have fired" events accumulate, a one-time toast offers to enable the nudge. The toast respects user preferences with explicit "not now" and "don't ask again" options.
The combined effect is auto-resume that escalates from "definitely safe" to "verified by a model call" before resuming, with discovery prompts that surface the feature without nagging.
The unreleased release — auto-open disabled by default
The most recent change in the changelog is a behavior reversal. Earlier versions auto-opened Claude Code session tabs on startup resume, after rate-limit recovery, and during the watchdog's stall recovery. The auto-open paths shipped buggy — tabs opened in the wrong workspace context, and the rate-limit recovery loop produced pop-the-tab-up-again loops that were hard to escape.
The fix in [Unreleased] is to gate all auto-open behavior behind a single dev-mode setting (powerClaude.devMode.autoOpenSessions, default false). Manual resume — the dashboard panel, the Session Explorer right-click, the command palette entry — all work exactly as before. What changed is that the extension no longer pops tabs at the user without their action.
This is the boring kind of changelog entry that does not get marketed, but it is the one that most improves day-to-day quality. Surprise tab pops were the second-most reported annoyance after rate-limit interruptions. Removing them by default makes the extension feel quieter.
What has shipped, in one paragraph
Over the past several months: the proxy core that produces pre-emptive rotation; parallel account distribution that scales throughput with pool size; account pinning and tagging that lets users shape rotation without code changes; the cost analytics dashboard reading from JSONL transcripts; the Session Explorer for navigating session history; the Insight Graph for AI-edited code attribution; confirm-gated auto-resume with discovery telemetry; the auto-open reversal that makes the extension quieter by default. The combined effect is an extension that does its work in the background and surfaces just enough information for the developer to know what is happening without demanding attention.
What is in flight
Public roadmap items that have been hinted at in the changelog and product page:
- Cap-hit recommender. Combining current usage with rolling-window state to estimate "you are this many turns away from the cap" and suggesting which accounts to favor.
- Subscription savings calculator. For pool users, an estimate of what the same workload would cost on the API at current rates. The dollar value of running on subscription instead of pay-as-you-go.
- Team Insight Graph. Cross-developer attribution for organizations where multiple engineers run Claude Code against shared codebases. The single-developer version of attribution is local-only; the team version requires a shared store.
- Per-project budget gates. Hard caps on dollar spend or token usage per project, enforced at the proxy layer.
Versions and timelines are not committed publicly. The public changelog records what has shipped; in-flight features ship when they are ready.
Frequently asked questions
Where can I see the full version-by-version changelog?
At neural-llm.com/changelog. Each version has its full release notes, ordered newest-first, with Added, Changed, Fixed, and Removed sections per release. The format follows the Keep a Changelog spec.
Are there breaking changes I need to know about across recent releases?
The auto-open default change in the unreleased version is the closest thing to a breaking change recently — sessions no longer auto-open tabs on rate-limit recovery unless the dev-mode setting is on. Users who relied on the old behavior need to explicitly opt back into it. Prior releases have been additive without breaking changes.
How often does Power Claude release a new version?
Every two to four weeks, roughly. The cadence is feature-driven rather than calendar-driven — releases ship when a meaningful improvement is ready, not on a fixed schedule. Patch releases for bug fixes can land between feature releases.
Will Power Claude auto-update from inside VS Code?
Yes. The VS Code Marketplace handles update delivery on its usual schedule. Users who want to pin a specific version can disable auto-update for the extension in VS Code's extension settings.
Is there an RSS or notification feed for new releases?
The changelog page has an RSS feed at /changelog?format=feed. The GitHub repo at github.com/neural-llm/power-claude also produces release notifications when watched.
Closing
The arc of the past few months has been from "rotation that catches rate limits after they happen" to "rotation that prevents the user from seeing rate limits at all," with everything else — session resilience, cost analytics, history navigation — built on that foundation. The current Power Claude is the productized version of that work, available today on the VS Code Marketplace.
To run the whole stack on a session of your own, the 14-day Premium Trial is $0 today and you can cancel anytime before day 15 — it layers parallel dispatch, confirm-gated auto-resume, and cost analytics on top of rotation. Not ready to add a card? Install Power Claude Free instead: the 7-day free trial gives you full Pro access, no credit card.