v1.3.1

Opus 5.5 support, the full 1M context window by default, local model servers in the pool, and
precise control over which model an API profile serves — plus the fixes that made those safe.

✨ New

• Claude Opus 5.5. Priced correctly ($4 in / $20 out, cache reads $0.20), in the model
  list, the GPT mapping and the /model picker. Claude Code 2.1.280 made Opus 5.5 its "opus"
  tier, and the picker follows it — no duplicate entry.
• Force 1M is now the default for the context window. cu code asks Claude Code for the
  full million on every session; before, it fell back to 200K behind the proxy and compacted
  several times as often. Anyone who already chose a setting keeps it. Auto is still there
  as the cautious option.
• Force model for API profiles. Every request to the profile uses exactly this model,
  whatever Claude Code asked for. Default model — now also visible and editable after a
  profile is created — becomes its fallback, used only if the endpoint refuses the forced one.
• Local model servers can join the pool. A Base URL may be http:// when it points at your
  own machine or LAN (localhost, 127.0.0.1, 192.168.x.x, 10.x.x.x), so an MLX /
  LM Studio / llama.cpp server works without a certificate in front of it. Anything beyond your
  network still requires https://, and Codex profiles stay https://-only.
• Orchestrator on one account, workers on another. In a cu code --profile <account>
  session, subagents now go to the profile marked Forced in subagents — say, a Claude
  orchestrator with its workers on a local coder model. Both sides are held strictly: nothing is
  rerouted if one of them cannot serve.
• macOS HUD: starts always on top, Liquid Glass and Large on a first launch; API-key
  accounts show their estimated total cost, with the token cap as the ring.

🛠 Fixed

• "Waiting for API response · will retry … check your network" on long turns. A request
  whose model took more than three minutes to start answering (a big Opus turn, a slow local
  prefill) was abandoned and re-sent by Claude Code while the first was still running. The
  proxy now answers slow streaming requests at once and keeps the connection alive with pings
  until the model responds.
• Codex reasoning effort "ultra" always failed (#8).
  OpenAI no longer accepts it, so it is gone from the list, and a setting saved with it now
  means "max". When a model refuses an effort that others accept, the request is retried once
  with the nearest effort that model supports — instead of switching to a different model.
• "Prompt too long" was mistaken for "model not found". A small local model rejecting an
  oversized conversation triggered a retry with a different model, logged a misleading
  "unavailable" line, and could leave the session stuck. The real error now reaches Claude Code,
  which compacts and carries on.
• A refused model was re-tried on every request. The refusal is now remembered per profile,
  so the next request goes straight to the model that works. Gateways that serve many models are
  unaffected.
• "Test" reported a working API endpoint as broken by probing it with a Claude model it does
  not serve. It now tests the model the profile is configured for.
• A profile could be accepted and then refused when used — the rule for a valid Base URL
  lived in two places that disagreed.
• Fable 5.1 and Mythos 5.1 cache reads were priced at 4× their real rate when the live model
  catalogue was unavailable.
• HUD: returns to the display it was left on after a relaunch; "Reauth" no longer cut off
  in a horizontal dock; a weekly bar's colour now matches its figure.

⬆️ Install or update

Already installed? It updates itself (or Settings → Updates → Check now).

curl -fsSL https://raw.githubusercontent.com/DevDock-AI/claude-unlimited/main/install.sh | bash

_{HUD-1.3.1-macos.zip below is the macOS HUD (universal, macOS 13+). You don't need to
download it — Claude Unlimited installs and verifies it for you.}

v1.3.0

The biggest release yet: a heads-up display for your pool, a full statistics view, the 1M
context window back in cu code, smarter handovers when one model runs out, and a long list
of fixes that make ChatGPT/Codex accounts and the dashboard far more solid.

✨ New

• HUD — Heads-Up Display (macOS). A floating dock showing every account at a glance.
  It installs and updates itself, and starts when you log in.
  • Two rings per account: the 5-hour window outside, the weekly one inside. Same colours
    as the dashboard, going red at the same points.
  • A red Reauth when an account needs signing in again. The logo takes the colour you
    gave the account.
  • Snaps to any screen edge and docks there. Dark, Light, or Liquid Glass on macOS 26.
  • Right-click to take over an account, turn it on or off, or choose which window the
    percentage shows.
  • Hover for a detail card: plan, windows, reset times, today's spend, and what it is
    serving right now.
• Statistics. Spend and tokens over time, broken down by model, project and account. It
  also shows what each request asked for next to what served it, plus an activity heatmap.
• The 1M context window through the pool. Claude Code falls back to 200K behind any
  proxy. cu code now gets the full million back on Claude accounts. A ChatGPT account in
  the pool is checked turn by turn, so a mixed pool keeps the million too.
• Per-model limits. When your plan's Fable week runs out before the account does, the
  session can move to another account instead of failing every Fable request. This works
  for Claude and ChatGPT accounts.
• ECO (off by default). Shortens tool output before it is sent, so the same work costs
  fewer tokens. Only tool output is ever changed, and every error is kept.
• Primitive speech (off by default). Asks for shorter replies, and Statistics measures
  how much it saves.
• Go back to your first-choice account once it has refilled (off by default). It only
  switches while the pool is idle.
• ChatGPT credits. Your balance is always shown. Spending credits once the plan window
  is used up is opt-in.
• Your own apps get sessions. Send x-claude-unlimited-session and a batch of workers
  spreads across accounts, each keeping its cache warm. Take over now holds until you
  release it.
• A clear answer when everything is spent: "all profiles are out of capacity — the
  first one resets at HH:MM", said once rather than on every request.
• Local SQLite store for usage, activity and projects. Your old logs are imported
  automatically on first start.

🛠 Fixed

• ChatGPT/Codex accounts
  • Rotating onto a Codex account no longer drops the connection mid-turn.
  • Streams that break now end cleanly instead of leaving an agent waiting forever.
  • Failures are reported as errors, not as empty replies.
  • An empty text block no longer breaks every later request in a session. Sessions
    already affected are repaired.
  • Auto mode's safety check works on Codex accounts.
  • Codex keeps a stable cache identity per conversation, and reasoning is carried across
    turns: about half the uncached input on a typical tool loop.
  • Cached tokens are now counted.
  • Two turns refreshing the same Codex login at once can no longer knock the account out.
• Dashboard speed
  • Leaked database handles are gone: pages that took 10–20 s now answer in under 100 ms.
  • Idle CPU went from ~22 % to 0 %.
• Subagents are recognised correctly, and nested ones tell Codex who spawned them.
• Re-authenticating an account now reads its usage straight away.
• ECO's setting was never read back, so ECO never actually ran. It does now.
• Statistics no longer shows your full local folder paths as project names.
• HUD
  • It returns to the monitor you left it on.
  • cu hud remove now stays removed.
  • A failed download is retried in the background.
  • It runs on Intel Macs as well as Apple Silicon.
• Windows
  • Service logs are no longer empty.
  • The claude client is found when the service runs with a minimal PATH.
  • Several tests that failed only on Windows are fixed.

🔒 Release & quality

• HUD builds are verified against a checksum shipped in the source tree before anything
  is installed.
• A release is published only once its HUD is attached and verified.
• CI runs the suite on macOS and Linux (Python 3.10 and 3.12).
• README and in-app Help are updated in all four languages.

⬆️ Install or update

Already installed? The app updates itself (or use Settings → Updates → Check now).
On macOS the HUD arrives on its own after the update — nothing to download by hand.

New install — macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/DevDock-AI/claude-unlimited/main/install.sh | bash

New install — Windows (PowerShell):

irm https://raw.githubusercontent.com/DevDock-AI/claude-unlimited/main/install.ps1 | iex

Then run cu code.

_{HUD-1.3.0-macos.zip below is the macOS HUD (universal: Apple Silicon and Intel, macOS 13+).
You don't need to download it — Claude Unlimited installs it and verifies it against a checksum
shipped in the source first.}

v1.2.7

Changes since v1.2.6

• chore(release): 1.2.7
• feat(models): editable parity list, effort on both sides, cost-ordered picker; fix cu launcher, aca alias, and update-restart
• docs(cli): document the cu alias in help, README, and dashboard Help

---

Install or update:

curl -fsSL https://raw.githubusercontent.com/DevDock-AI/claude-unlimited/main/install.sh | bash

v1.2.6

Changes since v1.2.5

• chore(release): 1.2.6
• feat(cli): show | parity in /model; fix stuck "Used now"
• feat(models): source model names and pricing from the LiteLLM catalogue

---

Install or update:

curl -fsSL https://raw.githubusercontent.com/DevDock-AI/claude-unlimited/main/install.sh | bash

v1.2.5

Changes since v1.2.4

• chore(release): 1.2.5
• fix(auth): refresh OAuth tokens with Claude Code's exact request shape

---

Install or update:

curl -fsSL https://raw.githubusercontent.com/DevDock-AI/claude-unlimited/main/install.sh | bash

v1.2.4

Changes since v1.2.3

• chore(release): 1.2.4
• fix(windows): stop console windows flashing once a second
• docs(readme): state the platform-support reality plainly
• fix(auth): persist the refresh backoff across restarts
• docs(readme): clearer per-platform install one-liners
• feat(desktop,windows): support claude-unlimited desktop on Windows
• fix(auth): tell the truth about where this OS keeps the Claude Code login
• fix(install,windows): make install.ps1 work on a bare Windows machine

---

Install or update:

curl -fsSL https://raw.githubusercontent.com/DevDock-AI/claude-unlimited/main/install.sh | bash

v1.2.3

Changes since v1.2.2

• Add a one-line Windows installer (install.ps1); v1.2.3

---

Install or update:

curl -fsSL https://raw.githubusercontent.com/DevDock-AI/claude-unlimited/main/install.sh | bash

v1.2.2

Changes since v1.2.1

• chore(release): 1.2.2
• fix(windows,linux): make the daemon actually run on Windows and Linux

---

Install or update:

curl -fsSL https://raw.githubusercontent.com/DevDock-AI/claude-unlimited/main/install.sh | bash

v1.2.1

Changes since v1.2.0

• chore(release): 1.2.1
• fix(auth): stop a rate-limited account from getting stuck needing re-auth

---

Install or update:

curl -fsSL https://raw.githubusercontent.com/DevDock-AI/claude-unlimited/main/install.sh | bash

v1.2.0

Changes since v1.1.8

• chore(release): 1.2.0
• docs: bring the docs back in line with what ships
• fix(storage): never let bookkeeping break a real request
• fix(config): validate a Profile field before it reaches disk
• fix(rotation): stop transient failures stranding a healthy account
• feat(dashboard): give every view its own URL
• feat(cli): route the Claude desktop app through the pool
• feat(codex): name the wire format a Profile speaks
• feat(codex): choose the GPT model and reasoning effort per Claude model

---

Install or update:

curl -fsSL https://raw.githubusercontent.com/DevDock-AI/claude-unlimited/main/install.sh | bash