What's new

v1.8.1 - 30th October 2026

🛠️ Bug Fixes

• Mistral models now accept requests for a deterministic answer. Asking a Mistral model to give the same answer every time (for example, by setting the temperature to 0) was being refused before the model even saw the prompt, with an unhelpful "top_p must be 1" error. The proxy now tidies up the request in the background so it always reaches Mistral in the form it expects. You do not need to change anything in how you use the app.
• The fix applies to every Mistral model, not just the one where the problem was first noticed. Ordinary requests, where some variety in the answer is welcome, are sent exactly as before.

---

v1.8.0 - 21st October 2026

🎉 New Features

• Run more than one AI application at a time: Visual Studio 2026 on one model and VS Code on another can now work through the app together, with each application's traffic kept separate from the others.
• A new Agents in flight section on the Overview page shows each busy application as it happens, including what it is doing and which model its work is on. Select an application's tile to see its full share of the traffic.
• Model slots on the Overview: the last model you used stays on screen, with any other models in use beside it.
• Spending by application: the Stats page now breaks usage down by application, alongside the existing model, provider and key breakdowns.
• Optional fair limits: if one application tends to flood the app, you can give each connection a sensible allowance so no single application crowds out the rest.

💡 Improvements

• The Agents in flight section can be folded away: a small chevron beside its heading tucks it up so only the heading remains, and the app remembers your choice the next time it opens.
• The Overview page keeps its shape: nothing on it grows or shrinks as models and applications come and go, so the rest of the page never shifts under you.
• A clearer name: the link that opens the model list is now labelled View Models.
• Fewer retry pop-ups: related notifications within a short window are now grouped into one, so a busy moment no longer floods your screen.
• The request log and the statistics can now show which application sent each request, with a sensible label given to any application that does not introduce itself. Nothing needs configuring.

🛠️ Bug Fixes

• Applications no longer interfere with each other's retries: two applications working at the same moment could previously overwrite one another's retry attempts.
• Retries no longer arrive all at once: the wait before a retry is now varied, so several applications cannot retry at the same instant and overwhelm a provider together.
• The Overview no longer grows taller as extra models appear: the row of models now scrolls sideways and keeps to one line, with the full model name shown on hover.

---

v1.7.6 - 10th October 2026

🎉 New Features

• Thinking levels that match each model: the thinking picker now shows only the effort levels each model actually supports.
• Cheaper repeat prompts: when a provider discounts a prompt it has already seen, that discount is now reflected in your costs.

💡 Improvements

• A clearer warning for an unknown model name: the app says the model is not recognised, suggests the correct spelling and still lets you continue.
• Chart zooming on the Stats page now happens only while Ctrl is held, so scrolling with the mouse wheel no longer zooms by accident.
• The spend summary wording now matches what the counters actually show.

🛠️ Bug Fixes

• Copied model names now work: a name copied from the app's own model list was being rejected as invalid. It now routes correctly every time.
• Budget spending now appears for providers with no spending limit: it previously stayed at $0.00, hiding what those providers had spent.
• Some model prices were out of date and have been refreshed to match each provider's current rates.
• Requests with no usage figures are no longer reported as using a model with no price.
• Token limits are now described as varying by model, rather than quoting one model's figure for a whole family.

---

v1.7.5 - 29th September 2026

🎉 New Features

• Model names with a slash can now be typed exactly as the provider publishes them, and are passed through untouched.

💡 Improvements

• Smarter model names: the app looks names up in the provider's published list instead of guessing, so every request carries the exact published spelling.
• The help now explains that some provider names contain hyphens and shows both ways to write them.

🛠️ Bug Fixes

• Fixed 59 models from providers whose names contain a hyphen - including every Z.ai GLM, xAI Grok and Meta Llama model - which could not be used at all before this release.

---

v1.7.4 - 29th September 2026

🎉 New Features

• Changing the proxy port now really works: a port saved in Settings is stored and used from the next start. Previously the change was quietly ignored.
• Restart now: after changing the port, the app offers to restart itself so the change takes effect.
• The Overview shows the port the service is actually running on.
• If a port is already taken, the message now names the port that is in use.

💡 Improvements

• Your API keys are never read back in plain text when settings are loaded or saved - they are masked.
• The proxy now only accepts connections from your own computer, rather than anything on your network.
• The help walks through the four steps of moving a port, in the order that works.

🛠️ Bug Fixes

• A port typed into Settings is now saved even if you press Save before leaving the box.
• The notice after a port change no longer claims the service restarted when nothing had restarted.

---

v1.7.3 - 28th September 2026

🎉 New Features

• Peak and off-peak pricing: models that cost double at busy times now show it, and prices are labelled with the peak multiplier.
• A one-off welcome message on first run asks you to accept the terms and privacy policy before continuing.
• A Support card on the Help page opens the official support site in your browser.
• A setup guide card on the Overview links to the setup guide and Help - close it once and it stays away.
• The crash log on the Help page can now be selected and copied.

💡 Improvements

• An estimate note now appears on every page that shows spending figures, so it is clear the costs are estimates.
• The help explains in more detail why recorded spend can differ from your provider's own bill.

🛠️ Bug Fixes

• Daily spend was overstated by roughly seven times for some DeepSeek models because repeat-prompt prices were far too high. Costs are corrected and your history has been recalculated.
• Fixed a crash at launch caused by the new estimate note.
• A DeepSeek vision model that was missing a price can be costed again.

---

v1.7.1 / v1.7.2 - 21st September 2026

🎉 New Features

• Reasoning cache figures are back on the budget card: hits, misses and hit rate since the app started.

💡 Improvements

• Budget thresholds now move in 25 cent steps instead of 50, so a $4.00 cap can use a $3.75 threshold.
• The help example and captions now match the new 25 cent steps.

🛠️ Bug Fixes

• The Remove button on budget rows is now lined up with the controls beside it.
• The budget card's refused-tier note now shows the right figure for the new steps.

---

v1.7.0 - 20th September 2026

🎉 New Features

• A redesigned Overview: cards for the service, the model in use, budget, cache and session continuity, with banners that no longer jump the layout when closed.
• A budget ring: a donut chart fills as the day's spending grows and closes when the cap is reached.
• Cache activity counters showing what has been stored and evicted since the app started.
• A rebuilt Help page: a hub of topic cards with a breadcrumb trail and FAQs grouped by subject.
• Smoother animations: cards fade in on load, the status dot crossfades, and a progress ring appears while the service starts or stops.

💡 Improvements

• The cache card now shows activity and holds its own actions: toggle cache, clear cache, clear logs and refresh.
• Summary tiles renamed for clarity - In cache now says exactly what it means.

🛠️ Bug Fixes

• Overview cache figures that could never change have been replaced by counters that update correctly.

---

v1.6.0 - 18th September 2026

🎉 New Features

• A cost ladder: as the day's spending crosses each threshold, the budget steps down through your ordered list of cheaper models instead of simply stopping at the cap.
• A rebuilt Budget page: add and remove steps, see which step is in use, and set a reserved amount.
• A warning when swapping to a cheaper model might cost more than it saves, because the cache built up for the old model is abandoned.

💡 Improvements

• All-or-nothing validation: one misplaced step explains why the whole ladder is refused, so the screen never accepts a ladder the proxy would refuse.
• Every model swap is recorded with the reason, and the Budget page shows which model is answering and why.
• Help and FAQs rewritten around the new ladder, including why my model was ignored.

---

v1.5.0 - 15th September 2026

🎉 New Features

• Usage history that survives restarts: token counts, timings and estimated costs are saved locally and kept for 90 days.
• A new Stats page: pick a range, break spending down by model or provider, view a spend chart, export to CSV and pop a chart out into its own window.
• Budget caps, alerts and hard stops: daily and per-provider limits with alerts when you cross a threshold, and a clear stop at the cap.
• Cache savings: see the money and time saved by cached prompts.

💡 Improvements

• Requests are sent in a consistent order so provider-side caching actually hits, giving faster and cheaper replies.
• Settings gained budget, usage-retention and cache sections, and configuration export now includes them.
• The Help page was rewritten.

---

v1.4.9 - 9th September 2026

🎉 New Features

• New models: DeepSeek vision and legacy models, Qwen 3.8 Max and Flash, Z.ai GLM-5.3 and GLM-5.3 Flash, and xAI Grok 4.6 and 4.20.
• More xAI models now accept every thinking level.

💡 Improvements

• The thinking cache saves as you chat and survives a restart.
• Failed streamed requests now log their full details in Session Logs, just like non-streamed ones.

🛠️ Bug Fixes

• DeepSeek thinking-mode chats no longer fail part-way through long conversations with a missing-thinking error.
• Images now reach the DeepSeek vision model - they were being dropped before the request was sent.

---

v1.4.8 - 11th August 2026

🎉 New Features

• A request inspector: open any entry in Session Logs to see the full request and response, with copy buttons and an export-to-file button.

💡 Improvements

• Streamed responses are stitched back together so you can read them in full, exactly as the client received them.
• Clear notices when a body was not captured or was too large.

🛠️ Bug Fixes

• The log detail panel now actually shows the request and response bodies - the sections were missing.