Changelog

What's new in Mel, release by release.

Everything that has shipped in Mel, release by release, newest first. During the beta, releases move fast — check back here (or the in-app About settings) to see what changed. For how updates reach you, see Updating Mel.

v0.1.45 — September 2026 (In progress — not yet released)

The agent can no longer tell you it did something it did not do. A turn that says “Done — I updated the file” without having touched anything is refused, as is one that promises to act and then just talks; after repeated promises with no action the turn ends with a visible note saying nothing was changed, instead of finishing quietly as though it had worked.
A long job runs to the end by itself. Work used to stop partway with a “tool-call budget reached — press Keep going” message; the same boundary is still there for housekeeping, but Mel crosses it on its own and says nothing, and Stop still works at any moment.
Answering the agent with “yes” no longer runs the Unix yes command and fills the screen with the letter y. Short replies — yes, no, ok, sure, continue — are read as answers; put ! in front to force a shell command.
A command that prints endlessly can no longer use up all your memory and take Mel down with it. Output waiting to be drawn is now bounded, the command is never slowed down, and if anything is dropped Mel says how much.
Asking to read a file gives the whole file. Longer files were being cut in the middle as they arrived, so the agent kept re-reading the same file and getting the same cut.
The agent knows where your conversations are kept and how to reopen them, instead of saying no transcript store exists.
A run_command written across several lines no longer paints over the conversation above and below it, and a heredoc whose end marker never closes is refused before it can write itself into your file.
The share icon in the conversation bar opens Export — it was drawn but did nothing. A wide table's scrollbar no longer covers its last row, and the working-directory chip shortens instead of running under the context gauge.
Plain-English build requests reach a model that can build them: “I want something that reads the emails and writes the reply by itself” was being answered by the smallest, cheapest model. A model that stalls, goes silent or answers nothing is set aside for a while instead of keeping the rest of a long turn.
Claude 5 Sonnet, when picked by hand, no longer spends its whole answer thinking and returns nothing.
If Mel stops being able to save a conversation because the signed-in account changed, it now says so instead of staying quiet.

v0.1.44 — September 2026 (Latest)

Eight connectors, each with YOUR OWN key, under Settings ▸ Connectors: GitHub (repositories, branches, commits, files, issues, pull requests incl. diff / create / merge, Actions runs, releases, code and issue search — 25 tools), Vercel (projects, deployments, build logs, redeploy / promote / cancel, environment variables with values never read back, domains, aliases — 19 tools), Alpha Vantage (quotes, price history, fundamentals, news sentiment, forex, crypto, indicators, economic and commodity series — 16 tools), Stripe (balance, customers, charges, payment intents, subscriptions, invoices, products and prices, payment links, refunds, payouts, disputes, events — 24 tools), Shopify (your store's products, orders, customers and collections — 10 tools; paste the Dev Dashboard app's Client ID and Secret with the store domain and Mel mints the token), Supabase (Supabase's own MCP tools — SQL, tables, migrations, logs, edge functions, branches — with your personal access token), Zapier (whatever actions your Zapier MCP server enables; the list is read live) and ElevenLabs. Your key is stored on this computer and sent only for calls you approve; a connector never runs on Mel's account.
Connector results are drawn as cards: repository and issue lists with state pills, pull requests with their branches and diff, Vercel deployments with build state and logs, Alpha Vantage quotes with a sparkline for every price series, Stripe balances, customers and prices, Shopify products and orders — every amount with its currency stated. Every link in a card opens in your browser.
A connector's tool list is no longer cut in the middle when it is long. With 24 tools the model was shown about half the list and told users that tools which existed did not; the whole list now reaches it, and a list too long to show whole says how many tools were left out.
Fourteen models are withdrawn from the picker for now: the eight most expensive OpenAI tiers (o1-pro, the GPT-5 Pro line, GPT-4 and GPT-4 Turbo, o1) and the Claude Opus 4.x line, Claude Opus 5 and GPT-5.6 Sol. A single short prompt on the dearest of them could bill the equivalent of a whole $10 pack in one reply; the worst case is now a twentieth of that. They come back when the per-reply cap is priced properly.
Every connector call still shows an approval card first — Mel always asks before a connector is used.

v0.1.43 — September 2026

The agent conversation uses the width you give it: the transcript column is centred and grows to 1280 pt (was a left-anchored 900), and the answer's prose has no fixed measure of its own any more.
Answers are set in the proportional face; code blocks and inline code stay monospace, so prose and code read differently again.
One card chrome everywhere — rounded 10 pt, hairline, soft shadow — for tool runs, shell output, diffs, tasks, approvals, questions, PDFs and audio. Disclosure chevrons rotate instead of swapping glyphs; card headers light up on hover; a running tool row spins.
The approval and question cards are neutral cards with an amber bar on the left (the approval's breathes while it waits), the command sits in a code well, and the buttons are a filled primary and a ghost secondary.
Your own turn is a soft accent-tinted bubble; new items fade in; three pulsing dots show Mel waiting on the model before the first token.
The composer is a rounded card whose outline warms to blue while it has focus, with a compact chip row, a real model chip carrying the maker's mark, and painted attach / dictate / send buttons.
New sessions open in the Mel agent by default (Settings ▸ Features turns it off; Esc still drops to the terminal).
Tool, shell, diff and task cards — and the thinking card — settle and fold about 1.5 s after they finish; click any header to expand it again. Answers stream in smoothly instead of appending in jumps.
Real maker marks in the model picker for Claude, GPT, Gemini, DeepSeek, Llama, Kimi, Qwen, GLM, Grok, MiniMax and NVIDIA; painted like / dislike / copy on every answer; a new Melling… indicator with a turning spark instead of a moving line.
The question card, the usage panel (tokens · cache hit · context · time as tiles), the Tools ▸ Preview panel and plan mode were redesigned and tested against live turns.
One chrome for every menu, list, panel and card: the + new-session menu, the account menu, notifications, the conversation's ⋮ menu, the Tools panel (explorer, conversations, search), agent runs, cloud jobs, and the export and continue-on-your-phone cards. Escape on a card now closes only the card.
Settings rebuilt: a rail with a glyph for every page in three groups, grouped rows on every page, the content column spans the whole window, theme as a Dark / Light / System control, the nine most-used shortcuts as key chips, a Privacy page that lists what stays on your machine, and an About page with Check for updates.
Dark theme: the neutral surfaces moved from pure grey to a blue-tinted neutral of the same luminance. Light theme unchanged.

v0.1.42 — September 2026

Long tasks no longer die with “the relay stream ended without a result (connection dropped)” about 30 seconds into a step where the model is quiet — the relay now keeps the connection alive while it waits on the provider, and the edge timeouts in front of it were raised. This was a server-side fix that reached every installed version on 18 September.
The per-tier output cap introduced for shorter answers is gone: at 1024 tokens it also cut off every file the agent tried to write on a normal coding turn (“cut off at Mel's 1024-token output limit”). One 16,384-token ceiling again; answer length is handled by the prompt.
When a long build turn reaches its round budget with changes it never verified, the agent now gets one round to run a build, typecheck or test before summarizing — instead of ending with “run these commands yourself”.
The agent researches or crawls first, writes a short numbered plan and confirms it once, then works through it without stopping when asked to build or clone something substantial.
macOS: the download is now a signed, notarized .dmg — open it and drag Mel to Applications. The .zip is still published for the in-app updater and as a portable download.
Drag a file anywhere over Mel (Windows, macOS, Linux/X11) and the whole window dims with an Add anything drop target — images, PDFs, and other attachable files go into the open chat; otherwise the path lands in the prompt you can see. (Pure Wayland still has no OS file-drop events from the window toolkit — use + or paste there.)
Paste a screenshot with Ctrl+V (Cmd+V on macOS) in agent chat attaches the image on Windows, macOS, and Linux (X11 and Wayland) instead of Paste failed — the clipboard could not be read (supported mime-type is not found).
Drag-select in agent chat keeps scrolling when the cursor hits the top or bottom of the transcript — selection is no longer limited to what is on screen (Windows, macOS, Linux).
Vercel, GitHub, Netlify, Railway, Fly, Render, and other common deploy hosts open on click without the “open this link?” toast (unknown sites still ask).
When Mel finishes create_pdf, the PDF opens in your system viewer right away — no extra click on the path.
When a localhost / loopback URL appears in the chat, Mel opens Tools ▸ Preview and shows a Preview chip with the host so the running app is one click away (Warp-style live preview).
The latest tool-run group in agent chat stays expanded after the tools finish (older groups still collapse).
Shorter agent answers on simple questions — system prompt length rules plus tiered max completion tokens (Trivial 512 → HugeContext 16k).
mel -p /path and mel --path /path open Mel in that directory (same as the positional mel /path).

v0.1.41 — September 2026

DeepSeek Fast slot is now deepseek v4.1 flash (live id deepseek-flash, native vision). Auto routing bias unchanged.
ElevenLabs Mel Connect TTS defaults to a free-tier-safe premade voice (Roger); library voices still work on paid plans when you pass voice_id.
Hosted relay and download CDN cut over to the new AWS account edge (CloudFront → US East).

v0.1.40 — September 2026

Tab accepts ghost suggestions in the terminal and agent prompt; the code editor prefers LSP completion (local identifiers as fallback).
Tools ▸ Preview shows the last localhost URL and browser screenshot beside the chat.
☁ in the agent footer runs a Mel cloud task; the titlebar Jobs chip lists open cloud jobs.
Share a conversation as a public read-only link (secrets scrubbed; 30-day expiry).
Settings ▸ Account ▸ Cloud triggers: mint meltrg_ tokens and fire POST /v1/triggers/run from GitHub Actions / cron.
semantic_search tool — local TF-IDF over the project (no paid API).
First-run landing shows six starter prompts when you have no recent conversations.
Mel Connect MCP gateways: GitHub, Slack, Supabase, Stripe, Shopify, ElevenLabs, Zapier, Vercel, Alpha Vantage (pass-through; publish still ops).

v0.1.39 — September 2026

Continue on my phone… opens and stays open — the click that opened the card no longer dismisses it on the same frame (Export had the same race).
The GitHub App install button from a failed handoff goes to a working install page (the old permissions URL 404'd).
Downloads go through CloudFront (global edge) instead of only Mumbai S3 — faster worldwide pulls; the updater follows.
Mel picks the closer of the India and US East relays at startup (health race); MEL_RELAY_URL still overrides.
Relay: session lookup is memoised for five minutes per process so consecutive legs of one turn stop re-paying the same DynamoDB GetItem; sign-out clears the memo.

v0.1.38 — September 2026

Long agent tasks keep going across Mel's ~90-second host legs — no fake "continued 0 times and reached its limit" stop, and no ask to keep going. The cut stays infra-only; Mel auto-continues the same turn (safety cap still 12).
When every model in the failover chain is unreachable, Mel drops the sticky lead for that turn so the next message can pick a live provider instead of re-hitting a dead one.
The "built, not verified" completion nudge fires at most once per turn, and a successful file read-back counts as verification — so Mel stops circling on the same downgrade notice.

v0.1.37 — September 2026

Mel stops re-reading the same file in circles. A repeated read now returns the bytes it asked for — it used to answer with a placeholder that said "re-read if needed", which is what sent the agent round the loop testers reported (up to 124 reads of one file, and no edit, in a single turn).
Long sessions keep their thread. Mel's repeat guard, its completion check and its empty-answer retry are now ON for everyone — a turn that says it finished without checking is told so and given one more step, a tool that has failed three times the same way is stopped instead of retried ten more, and an answer that comes back empty is named and retried once.
Mel stops saying "Task completed." when it did not. A turn that failed, ran out of credits, hit a rate limit, needed a sign-in, outgrew the model's context or simply died now says which of those it was — in the notification and in the transcript.
The agent's tools got a memory and a voice. A shell that remembers your cd between commands; reads that tell you which lines you got and how to ask for the rest; searches that say "found 250 of 3110" instead of quietly cutting the list.
Long conversations keep what you agreed. Mel decides once whether the chat is too big, measured against the window of the model that is actually answering — so a million-token model is no longer compacted as if it had 200k — and what survives is a checkpoint under fixed headings (decisions, plan, files) instead of quietly dropping the important bits first. Pinned models keep this relief too; auto before the first answer does not invent a window.
Something you tell Mel once stays true. A rule you state — use main, don't touch app/ — is offered on a chip in your own words; keep it and Mel checks it before file changes and refuses ones that would break it, quoting you back. The task checklist stops going dark and duplicating; Mel asks real questions with options and deadlines, and will not re-ask something you already answered. Pressing ▶▶ only skips approvals, not every question.
Mel's instructions to the model stop being one ever-growing block on every message. Rules now travel with the tools in use, so a one-line question no longer pays for browser advice Mel is not using — and a cut-off answer no longer invents a dead continue tool.
Long auto conversations no longer forget early: the context gauge and the compaction trigger now measure against the window of the model actually serving the conversation. Before, every auto conversation was measured against a flat 200k — on a 1M-window model that summarised-and-forgot at 16% of the window.
auto's failover chain is generated from the model catalog instead of hand-ranked lists — cheap and open-source providers lead by construction; when a provider's account runs dry Mel steps that whole provider out for ten minutes instead of trying its next model on the same empty wallet.
The Fable and Astra models have been removed from the model list. If you had pinned one, new conversations start on auto and Settings says which model was retired — instead of a chip naming a model that no longer exists and turns failing with "unknown model".
Mel keeps working until the credits you can see actually reach zero. If you had add-on credits while your monthly ones were spent, Mel was refusing turns against a hidden negative number instead of the two numbers on your Billing page.

v0.1.36 — September 2026

Connectors. Settings ▸ Connectors lists the partner tools Mel Connect offers — ElevenLabs first. Press Connect, sign in with your own account in your browser, and its tools appear to Mel behind the same permission cards as everything else. Your sign-in stays on this computer; disconnect any time.
A connector is asked about every time it is used. The card names the company and the exact tool — “ElevenLabs wants to: text_to_speech” — and says plainly that this sends your request to another company. There is no “always allow” for a connector, and a tick you gave a shell command earlier does not cover one.
Run your own tool server. Custom server… at the bottom of the Connectors page points Mel at a program on your own machine. Values you enter that look like keys go to your keychain, not into a settings file.
The model picker now lists every chat model the ten providers serve (143 ids, up from 61) with the providers’ true context windows — nine were too small (DeepSeek V4 is 1M, not 128k) — and the three ids no provider serves any more (Groq’s two Llama ids, Kimi K2.5) are gone. Picking a legacy id still works; auto only ever picks a model with a verified price and window.
The picker also lists NVIDIA’s NIM catalog (55 models: Nemotron 3.5 Lightning, Nemotron 3 Super / Ultra / Nano Omni and every other chat model the endpoint serves — Kimi K3, DeepSeek V4, MiniMax M3, Gemma, Llama, Mistral, Phi, Granite and more, marked (nvidia)). They can be pinned; auto never picks them yet — until Mel’s NVIDIA key and terms are in place a pinned NVIDIA model falls over to the usual chain with a visible “why this model” note.
auto reads the task instead of making one fused guess — what kind of work it is, how hard, how many tool rounds it will take, how big it is. A long conversation no longer makes a short question expensive: that one mistake was 75 of the 164 misroutes we measured. Mel’s own “keep going” continues on the previous turn’s terms, and a turn with an image in it is routed once, onto a model that can see it.
auto now tells every provider whether to think. A quick question or a small edit gets a no-thinking answer on every model that can be told to stop — DeepSeek, GLM, Kimi K2.6, MiniMax M3, Qwen, Gemini 2.5, the GPT-5.x line, Claude — while a build or a hard problem gets a small or medium thinking budget. Your own reasoning-level pick always wins. MiniMax models no longer show their reasoning inside the answer.
Bring-your-own-key users can now set an NVIDIA key.

v0.1.35 — September 2026

Approve from your phone. Link a Telegram chat in Settings ▸ Account and every permission request Mel raises also reaches your phone — a 200-character masked summary, never the full command or a file. Tap Approve or Deny; the first answer wins, the app says “Approved from your phone”, and there is no “always allow” from a phone by design. Off until you link it.
A reply from your phone becomes the next thing Mel does. Reply to one of Mel’s messages with a sentence — “use the staging database instead” — and when the current turn finishes, the app picks it up, shows it as “From your phone” and runs it. A message that is not a reply gets a short note back explaining that, rather than disappearing.
Mel can email you when a long turn finishes. Off by default; turn it on in Settings ▸ Account. A turn that ran for more than two minutes, or one you sent to the background, ends with an email containing the closing message and what the turn cost. Results only — Mel never asks you to approve anything by email, and each email threads onto the previous one for the same conversation.
Continue on my phone. In a conversation, choose Continue on my phone, pick how long — 30 minutes, 2 hours, or until a time you name, up to 8 hours — and see the price before you agree to it. Then close the laptop. Mel in the cloud picks that same conversation up, with its whole history, on a fresh clone of your repository, and you keep it going from Telegram: send it a message and it becomes the next thing Mel does, questions and permission requests come to your phone, and the answer comes back there. Send stop, let the time run out, or press Take back on the laptop, and the conversation comes home with everything that happened while you were away. Cloud time costs 6 credits an hour, billed by the minute, on top of the usual model credits. Paid plans only, because the API keys Mel uses on this computer never leave it — and it is not your laptop still running with the lid shut, it is a copy of your repository running on our machines.
Export a conversation. ⋮ ▸ Export conversation writes one .zip with everything Mel saw and did in that conversation — prompts, replies, thoughts, every tool call and its full output, shell commands, diffs, approvals and who answered them — plus a per-turn table of what each turn cost, which model answered it and why it was chosen, and a README. You can tick two options before it writes: mask secret-looking strings, and replace absolute paths and your home folder; both are off by default and the README counts what they changed. It stays on your machine — nothing is uploaded, and sharing it is your choice. If Mel circled or forgot, this is the file to send us.
Erasing your account now also unlinks your phone. Before this, “erase everything” left a linked Telegram chat behind that could still answer approval requests. It is part of the erase now.
A rate limit is no longer reported as running out of credits. When the service sheds load, Mel used to tell you your plan was exhausted — two different problems with two different fixes, and only one of them was yours. It now says which one happened.
A crash report names what it matched. The support log records the classification of a panic alongside it, so a report you send is diagnosable without a reproduction.

v0.1.34 — August 2026

Mel finishes the job instead of stopping to ask. A build task — clone this site, scaffold this project — starts by reading: the page, the bundle, the schema. All of that is read-only, and Mel was budgeting it as if it were a research question, so it ran out of room mid-investigation, before it had written a single file, and asked you to press “keep going”. Plan mode made it certain: nothing can change while a plan is being written, so every planning turn was scored as pure research. Mel now recognises a build request for what it is and gets the room to finish, while a turn that has genuinely stopped making progress wraps up sooner than before.
Mel can see the pages it builds. Its headless browser could not render WebGL, so every map, chart, 3D scene and canvas app came back as a black rectangle — and Chrome reported that as a successful screenshot. Mel would check its own work, see nothing, and fall back to reading the page's text. It now renders those pages properly, and a screenshot that comes back blank is reported as a failed render rather than passed off as evidence.
The rating buttons appear once per answer, not after every step. A long build used to leave a row of thumbs after each intermediate line — “let me check the logs”, “let me verify it renders” — rating nothing that had finished.
A link Mel gives you is clickable even when it writes it as code. The address of the thing it just built for you — the local server it started — used to render as plain text you had to select and paste. A code span containing a command is still left alone.
A message you send while Mel is still working is now queued instead of thrown away. It used to disappear the moment you pressed enter — the text was taken out of the box and dropped, with nothing on screen to say so — while the placeholder promised you could steer the running agent. Your follow-up now waits, shows above the input, sends itself when the turn finishes, and comes back to the box if the turn ends any other way. It is never sent into a turn that failed.
The agent stops forgetting what you already decided. In a long session, when Mel compacts older turns to fit the model's context window, it now carries the task checklist, the questions you already answered, the approvals you already gave, and its own rolling summary. Before, all four were silently dropped — which is why a long build would re-ask a settled question and start re-planning work that was already underway.
Clicking “Thought for 2 seconds” actually opens the thinking again. It looked like a button, it listened for clicks, and it never received one, so once a turn ended the model's reasoning became unreachable — even though it was saved on disk the whole time.
File paths the agent writes are now links. When Mel tells you a report is saved at a path, you can click it. Paths that would run something — scripts, apps, installers — are deliberately not clickable, and a link whose text does not show where it goes is refused.
PDFs got the rest of the way. Asking for one no longer spends an extra round-trip discovering that Mel can make PDFs — it starts straight away — and the card now offers Download, which saves a copy into your Downloads folder without ever overwriting a file already there. You can also read the document beside the conversation instead of opening it in a separate app.
Drag files from Finder or Explorer onto Mel to attach them. They go through exactly the same size limits and image handling as the + button, and anything Mel cannot attach lands in the prompt as a path instead of vanishing.
A turn that stops at its tool-call budget now says so and offers to keep going, instead of suggesting you commit an unfinished build. Running low on credits no longer tells the agent to cut its work short — it prefers a cheaper model instead of leaving the job half done.
The new-session menu names what it starts — Agent and Terminal — and the ⌘T badge now sits on the row that key actually opens. If you had starred a session config as your default, the badge was on the wrong row.

v0.1.33 — August 2026

Sending an image no longer closes Mel. On Windows machines with Snapdragon processors, attaching any picture and pressing send killed the app instantly — every image, every time, while PDFs were fine. Mel was releasing the thumbnail's graphics memory in the middle of the same frame that was still drawing it, which the graphics driver treats as fatal and which no error handler can catch. Nothing was wrong with the pictures: the file size never mattered, because what reaches the graphics card is a 96-pixel thumbnail.
Three more ways Mel could close on you, all fixed. Closing a pane or a session that contained images, and changing or removing your profile photo, both released graphics memory the same unsafe way. Separately, searching a project whose files contain certain letters — Turkish capital İ, the Kelvin and Angstrom signs, capital sharp S — crashed the moment results were drawn, on every platform, because the highlighter measured its offsets against a lowercased copy of the line and then applied them to the original.
Long tasks finish instead of stopping silently. Any single step that ran past about two minutes was being cut off by the server's request limit — the reply was lost, no file was written, and the agent often looked like it had succeeded. Requests are now handed over cleanly before that limit and continue where they left off, so a long build or a large file write completes. This one is already live for everyone; it needed no update.
The agent's thinking panel no longer collapses into a one-character-wide strip down the right edge of the screen. Past the first few thousand characters of a long thought, every later block was laid out in whatever pixels were left over on a row that had already been filled — so it rendered one letter per line while the panel beside it sat empty. It stacks properly now and reads at full width.
Multi-agent runs tell the truth. A worker whose turn died used to report "IDLE (done)" with its last half-finished sentence, so the agent leading the work believed a file had been written when none had; it now reads FAILED, with the reason, and the worker's own card says so too. A message sent to a worker that is still busy is no longer confirmed as delivered when it was in fact dropped.
Reopening a multi-agent session keeps its workers. Closing Mel used to scatter them into unconnected sessions and quietly erase the record of which belonged together; the group is now kept, each worker links to its own conversation, and the ones that are no longer running say so plainly rather than pretending.
Crash reports name what they matched. When Mel recovers from a graphics-driver crash it now records which fault it recognised and where, instead of only that it restarted. No prompt text, credentials or file contents are ever written to that log.

v0.1.32 — August 2026

Long sessions no longer hit a wall. A malformed reply from one AI provider could get saved into a conversation, after which every model refused it forever with a misleading 'provider outage' message — while new sessions worked. The server now repairs such conversations automatically on every request (existing stuck conversations start working again with no action needed), the malformed shape can no longer be saved in the first place, and the error message for genuinely unserveable requests now tells the truth.
Conversations that outgrow a model's context window now compact automatically and continue — a quiet one-line notice instead of an error. Very long sessions are the point of Mel; now the seams show it.

v0.1.31 — August 2026

Recording you can see. While dictating, the mic becomes a stop square with a live input-level meter and an elapsed timer, and transcription shows its own distinct state — no more guessing whether Mel heard you. A muted microphone now shows a flatlined meter while you speak instead of failing silently at the end.
A graphics-driver crash no longer closes Mel. If the renderer dies mid-session, Mel writes a crash record, restarts itself in compatibility mode, and tells you what happened — instead of vanishing with your session.
Changing plans is now a switch, not a second subscription. Upgrading (or stepping down) opens a Stripe confirmation that modifies the subscription you already have, with proration handled by Stripe. Subscribed users see 'Switch to…' in Billing, and moving from Max to Business is labeled as the step down it is.
Beta updates channel. Settings ▸ Features ▸ 'Receive beta updates' opts this install into pre-release builds; everyone else stays on stable. This release is the first to ship through the new channel.
16-bit PNG attachments (like the Mel logo) are normalized before upload so every vision provider accepts them.

v0.1.30 — August 2026

Windows on ARM: Mel now draws with Direct3D 12 on ARM machines instead of OpenGL. On a Snapdragon X laptop the OpenGL path repeatedly crashed the whole operating system under agent load; the same workload ran for eighteen minutes clean on Direct3D 12. The trade-off is occasional visual corruption under heavy load that repairs itself — an OS crash loses your work, a cosmetic glitch loses nothing.
If Direct3D 12 misbehaves on your machine, starting Mel with MEL_RENDERER=glow forces the previous renderer.

v0.1.29 — August 2026

Windows: the Direct3D 12 renderer was never actually compiled into any shipped Windows build. Mel asked the graphics library for a backend that was not in the binary, got an empty set back, and could not create a window — this, not a driver, is why v0.1.27 would not open on Windows at all. It is compiled in now, and a Windows CI gate fails the build if it ever falls out again.
A start-up failure now leads with the fix. The dialog opens with the recovery steps (start Mel with MEL_RENDERER=glow) rather than the diagnosis, and says plainly that reinstalling will not help — downloads are replaced in place, so a reinstall fetches the same build.
The support log records what happened to the fallback attempt, not only the first failure. Reading the old log needed the customer's own account of it to interpret.
Publishing a release no longer depends on GitHub. A failed GitHub-Release step used to skip publishing the download manifest, which left the download service holding new files under an old manifest.

v0.1.28 — August 2026

OpenGL is the default renderer again on every machine. v0.1.27 switched ARM Windows machines to a renderer that had never once been started on the hardware it was chosen for, and it did not open there.
A renderer that fails to start is no longer the end of the session. Mel retries once on OpenGL, and if that fails too it shows a native dialog naming the likely cause and where to find the support log. On Windows a release build has no console, so the previous behaviour was to exit silently with no window and no message at all.

v0.1.27 — August 2026

This release does not open on Windows — any Windows machine, ARM or Intel. It is superseded by v0.1.28 and v0.1.29; if you are on it, update. The cause and the guard against it are in the v0.1.29 notes above.
Long agent conversations stop getting slower as they grow. The transcript now lays out only what is on screen, using measured heights rather than estimates, so a conversation that has run all afternoon scrolls like a fresh one. On a 1.4 MB conversation the transcript's share of a frame fell from 8.37 ms to 0.40 ms and stopped growing.
The Thoughts panel no longer re-wraps its whole text about thirty times a second. A blinking cursor was rebuilding the entire reasoning body on every frame; it is now split into stable chunks, so watching the agent think costs what reading it costs.
Long command output says when it has been truncated instead of silently stopping at 64,000 bytes, and the agent's own long replies get the same visible marker.
Ending a turn no longer saves the conversation twice. It serialised the whole conversation two times in the frame that finished a turn.
Replies are bounded. On one afternoon a single account produced two agent replies of 65,536 tokens each, taking about twenty-one minutes apiece, against a normal median of 159 tokens in under five seconds — that runaway is what grew the conversation that made the machine struggle. Completions are now capped server-side, so this applies to every version.
A support log you can send us. Off-by-default logging did not help anyone who had already hit a problem, so Mel now keeps a rotating 256 KB log of its own events — starts, renderer choices, update checks, failures. It never contains your messages, prompts or command output, and it is deleted along with everything else when you erase your data.
About names the machine Mel is really running on, including when an Intel build is running under emulation on an ARM Windows machine. Support spent two days not knowing that.
The agent can read big pages again. Taking a snapshot of a page larger than about 64 KB used to hang for thirty seconds and come back empty; one real page was 1.18 MB. Screenshots also describe the picture rather than the request that made it.
There is no v0.1.26: its Windows build failed to compile and it was never published.

v0.1.25 — August 2026

Security: the library Mel uses to open links in your browser is updated for RUSTSEC-2026-0257, where a crafted BROWSER environment variable can inject arguments into the browser being launched. Mel deliberately loads your login shell's environment, which is where BROWSER comes from, so the path was reachable. v0.1.23 and v0.1.24 are affected equally — the advisory was published against a dependency that had not changed — so this is the release to be on.

v0.1.24 — August 2026

The Free plan moves to bring-your-own-key. Mel's AI agent on Free now runs on a provider key you add in Settings → AI: you pay your provider directly, at their price, with no Mel credits and no Mel markup. Groq and Gemini both offer free key tiers. Pro, Max and Business are unchanged and still need no key of your own.
If you already have a Free account, nothing changes for you — accounts created before this release keep their monthly hosted-AI credits.
Getting to your keys takes one step. Provider keys have their own entry in the avatar menu, instead of being the last section of the fifth settings page.
A rejected key now says so. If your provider key is expired or revoked, Mel tells you the key was rejected and where to fix it — it used to report a temporary provider outage and suggest waiting a minute, which never fixed it.
A key you supply can no longer affect anyone else. A failure on your own key is no longer counted against Mel's view of that provider's health.
Clearer limits. Free includes 3 agent web searches a day and 200 command suggestions a day; those run on Mel's keys, so your own key doesn't change them. Fair-use ceilings (currently 400 agent turns a day) apply on every plan, including with your own key — bring-your-own-key means no Mel credits, not no limits.

v0.1.21 — August 2026

Two accounts on one computer no longer share one history. Conversations are stored per account, so signing in as someone else shows you their conversations and not the previous account's — your existing conversations are moved into place once, automatically.
Every action answers. A whole class of buttons used to finish in silence — "Check for updates" ran a real network check and changed nothing on screen, and several messages Mel wrote had no place to appear at all. Actions now confirm, and the ones that fail say so.
Menus and pickers stop opening in the wrong place. Nine dropdowns were positioned from a height each caller worked out by hand, and four of them got it wrong, so lists could open clipped or off the screen.
The agent can supervise real work, not just write it: start a long build or a dev server in the background, wait for it to be ready, call it over HTTP, open the page in a headless browser to see what actually rendered, and look at images. It can also plan first and show you the plan before touching anything.
A conversation that had become unusable can be used again — one bad message could make every following turn fail, and Mel now repairs the stored history instead of failing on it.
When a model has a bad minute, the conversation no longer stays stuck on the stand-in that covered for it; the next turn goes back to the best available model.
Slow frames are now measured rather than guessed at, and the operations behind the worst freezes are barred from the code that draws the window.
If you're on v0.1.18 or later, this update is one click: avatar menu → Update and relaunch Mel.

v0.1.20 — July 2026

macOS gets a real application menu — Mel, File, Edit, View, Find, Window, Help — with the shortcuts you expect (⌘, for Settings, ⌘N, ⌘W). Accelerators are deliberately held back while a CLI coding agent has the keyboard, so ⌘P and ⌘D can't be stolen out from under it.
Window controls behave: on macOS Mel no longer paints its own minimize/maximize/close over the ones the system already draws — the traffic lights were always the real controls, and in fullscreen Mel's copies did nothing at all. Minimize, Zoom and Toggle Full Screen now live in the Window menu.
The macOS app icon is sized to Apple's grid, so it matches its neighbours in the Dock instead of towering over them.
Ticks, crosses and close buttons render on macOS. They were empty boxes — the system font Mel relied on skips that block of symbols, so Mel now ships its own.
The keyboard hints above the prompt are clickable, not just decorative — and two of them that advertised shortcuts Mel didn't have are gone.
The working-directory chip opens a directory picker, in the terminal and in an agent conversation: jump to a parent or subfolder without typing cd.
Every titlebar button now tells one story — the icon, the tooltip and what actually happens finally agree.
The starter session templates no longer depend on a third-party CLI being installed; they open Mel's own agent. The one that asked for a repository now explains itself and offers a way to add one instead of dead-ending.
PDF export works on a stock machine. If no browser-based renderer is installed, Mel now falls back to one it ships with, so "create a PDF" no longer fails with a missing-backend error.
Smaller repairs: the last session's close button no longer offers to do something you can't see; a failed command is visible in the sidebar and one keystroke from being handed to the agent; the sidebar's View-as and Density settings do something you can actually notice.
Not changed: the in-app Feedback link still opens the GitHub repository. Where it should point is still being decided, so it was deliberately left alone rather than guessed at.
If you're on v0.1.18 or later, this update is one click: avatar menu → Update and relaunch Mel.

v0.1.19 — July 2026

A new look: Mel's brand is now a true blue, with pink marking everything the agent does — the /agent label, the agent sparkles, and the running "Melling…" line. Every color pairing is contrast-checked so the app stays crisp and legible.
Bigger, more comfortable text by default — and a new Settings → Appearance → Font size control (13–20pt) so you can tune the whole app, terminal output included, to your eyes.
A cleaner app icon: the same M mark the website uses, everywhere.
If you're on v0.1.18, this update is one click: avatar menu → Update and relaunch Mel.

v0.1.18 — July 2026

One-click updates, for real: earlier builds could never see a new release (their update check pointed at a source that stopped answering) — Mel now checks a public release manifest, so the update prompt actually appears when a new version ships.
One-time step: if you're on v0.1.17 or earlier, download this version once from the download page — those builds can't update themselves to it. From v0.1.18 on, updating is one click inside Mel.
macOS: after an update Mel now closes and reopens by itself (it used to quit and wait for you to find it in the Dock), and updates survive the move across disk volumes.
Windows: updating no longer opens the full installer wizard — Mel closes, updates silently, and relaunches on its own.
A small dot on the avatar now signals when an update is ready, and the account menu gained "Check for updates" — a failed check says so honestly instead of pretending you're up to date.

v0.1.17 — July 2026

macOS builds are now signed with an Apple Developer ID and notarized by Apple — Mel opens on a normal double-click with no Gatekeeper "Open Anyway" step, and permission grants like Full Disk Access now persist across updates.
With Windows already code-signed, this is the first Mel release signed on both desktop platforms.

v0.1.13 — July 2026

Fixed upgrading from Windows: checkout links opened without their full address, landing on an error page — the browser now receives the complete link on every OS.
Windows builds are now code-signed (Concepts Dzign LLC).

v0.1.12 — July 2026

Your account now works on the web too — sign up or sign in at openmel.dev, see your plan, credits and usage at /account, and upgrade from the browser.
Sign in with browser: the app can sign in via a short code you confirm on openmel.dev — no password typed in the app.
New Profile and Security settings pages — edit your name, change your password, see active sessions and login history, and sign out other devices.
Plans in the app now show full feature lists with a monthly/annual toggle (annual saves 10%), matching the website.
Real in-app auto-update: Mel downloads, verifies (SHA-256) and applies updates itself, then relaunches.
A contact form at openmel.dev/contact, welcome emails on signup, and stronger rate limiting across sign-in and signup.

v0.1.11 — July 2026

Fixed a rendering bug where a long agent conversation could go blank when scrolled — the full response now stays visible from top to bottom at any window size.

v0.1.10 — July 2026

Cloud history sync — your agent conversations now sync to your account, so reinstalling Mel or moving to another machine restores your full history instead of starting over.
Multiple agent chats at once — open several agent conversations side by side in a new Agents sidebar; opening one from history adds it as its own chat instead of replacing the one you're in.
Agents finish the job — ask the agent to build something and it now runs the whole task to completion instead of stopping every few steps to ask permission. A one-key “Keep going” and a hands-off mode make long builds effortless.
Readable long conversations — very long agent transcripts stay within the window; wide terminal output, file paths, and reasoning are contained instead of being pushed off the side of the screen.
Running-turn controls — the buttons shown while the agent is working now have tooltips and actions: fast-forward the rest of a turn hands-off, or send it to the background and get notified when it's done.
Windows fixes — agent commands run correctly on Windows; closing a session or tab works on the first click; a not-installed command exits cleanly instead of hanging.

v0.1.9 — July 2026

Windows installer — a proper per-user setup.exe with desktop and Start-Menu shortcuts and a clean uninstall from Settings → Apps (the portable .zip is still available).
Readable agent responses everywhere — fixed invisible text and washed-out code cards in the agent transcript on light-themed systems; code blocks are now dark, horizontally scrollable, and copyable.
Sign-in reliability — fixed a sign-in re-prompt on Windows so your session persists across restarts.

v0.1.8 — July 2026

Credit plans — every plan now includes a monthly bucket of AI credits (1 credit = a cent of real AI cost): Free 150, Pro 1,500, Max 18,000, Business 1,500 per seat. Credits work across every frontier model; bring your own key on any plan and a turn costs 0 Mel credits.
Two new plans — Max ($200/mo) for power users who live in the agent, and Business ($50/seat/mo) with per-seat credits and priority support.
Add-on credit packs — 400/1k/3k/6.5k packs that roll over month to month and last a year; optional auto-reload with a monthly spend cap you control.
Self-serve billing — upgrade, buy credits, and manage your subscription from Settings → Billing; a new Usage tab shows your per-day credit history.
Account panel — sign up, sign in, and sign out directly from Settings → Account (no more detour through a conversation).
Image attachments — attach images to an agent message; image turns route to vision-capable models automatically.

v0.1.7 — June 2026

Self-serve accounts — the waitlist is gone. Sign up for a free Mel account instantly. The website now leads with Download / Get started everywhere.
Enterprise removed — Mel is a personal and small-team tool; the Enterprise/Teams tier and its self-host copy have been removed entirely.
Pricing simplified — down to two clean plans at the time (superseded by the credit plans in v0.1.8).
SEO and metadata — every page now has its own OpenGraph/Twitter card, sitemap, and robots.txt; optimised hero video.
Honesty fixes — removed claims for features that hadn't shipped yet; corrected slash-command names; removed premature open-source claims.

v0.1.6 — June 2026

Accounts & sign-in — added email/password Mel accounts with an in-app sign-in panel for the agent.
Account status panel in-app — a clear sign-in/account panel instead of an error when the agent needs authentication.
Website redesign — new ink-on-paper design system (Space Grotesk/Mono), real Mel screenshots, provider logos, and a self-playing hero demo video.

v0.1.5 — Earlier release

Agent orchestrator — /orchestrate drives worker Mel agents across panes with cross-agent shared memory.
Cross-provider fallback — when a provider is rate-limited or unavailable, the agent automatically reroutes to the next available provider and keeps going.
Refreshed catalog — 6 providers, 41 verified models.
Agent transcript scrolling + delete-conversation from the landing.

v0.1.1 — Earlier release

Live command streaming + Stop — watch long commands stream; Stop kills the whole process tree.
Faster orchestration — the lead delegates to workers earlier.

v0.1.0 — Initial public beta release

Multi-agent orchestration — /orchestrate drives worker CLI agents (Claude, Codex).
Agent over SSH — melt a remote session into an agent.
Built-in code review — diffs, inline comments, LSP.
Command palette + structured blocks.
Agent mode — 6 providers, many models, no API keys.