What's new in InteJer

No more "Session expired" about an hour in, and no more waiting a minute for the first reply

Signing in now lasts. After the hourly sign-in refresh, the agent kept sending the old, expired token for every model it had already used, so about an hour into a session you got "Session expired", and signing in again did not help until you restarted the app. New tokens now reach the agent straight away. If the sign-in service is briefly unavailable when the app refreshes, it now keeps your session and retries instead of giving up (before, one bad response could end the refreshing for good), and when a sign-in genuinely does end, the reason is written to the log.

The first message no longer waits for a broken tool. If the KiCad tool server could not start (for example because KiCad's Python packages are missing), it used to hold up any message you sent in the first minute after launch by up to about 50 seconds. It now fails immediately and gets out of the way.

Browser automation starts. The Playwright server was shipped without its files and failed on every launch. It now comes with them.

Startup no longer waits on a failed download. On slow networks the app could sit on "can't connect to Local Server" for 30 seconds or more while it tried to download a package version that does not exist.

Prompt optimizer, fixed. Optimizing a prompt no longer drops @file and @agent mentions, a second round of questions remembers the answers to the first, Retry keeps your answers, and a full quota or a sign-in problem now says so instead of a generic "try again".

More of the app in your language. The Updates panel, the prompt optimizer, the first-run tour, the Integrations settings and archived conversations are now translated in all 17 languages.

If you are on beta.17 or later, this update installs itself: InteJer downloads it in the background and offers "Install and restart".

The Model Device Standard, in the timeline

When the agent reaches a hardware tool, the timeline now says so in one language. A KiCad call no longer reads "Called kicad_route_trace"; it reads Model Device Standard · KiCAD · route trace, with the tool's logo. A PlatformIO build for an ESP32 board reads Model Device Standard · ESP32 · build, and once a session has named a chip, later calls that do not repeat it keep that name. Parts lookups read Model Device Standard · Parts.

Runs of calls to the same device fold into a single row, the way reads and searches already fold into "Explored": Model Device Standard · KiCAD · 14 operations, expandable to every call with its action and first argument, and marked when one of them failed. While a call is running the row says so.

Stitch, browser automation and any servers you add yourself keep their existing rows. Nothing changed under the hood: the same MCP servers run the same way, this is how their work is presented.

If you are on beta.17 or later, this update installs itself: InteJer downloads it in the background and offers "Install and restart".

Agent picker fixed, executors on GLM-5.3-Flash

The agent dropdown works again. In beta.18 the Build / Intejer-Agent / Intejer-Designer picker opened as a squashed strip with the names spilling out of it. It now opens as a normal list again.

A new executor model. Intejer-Agent, Frontend-Developer and Intejer-Designer now run on Z.ai's GLM-5.3-Flash instead of Gemini 3.7 Flash. It reads images, calls tools, and carries a 1.3 million token context, at a fraction of the previous per-token price. Gemini 3.7 Flash stays available in the model picker under OpenRouter, next to GLM-5.3-Flash, if you prefer it for a session.

Everything from beta.18 is in here too: the refreshed sign-in and onboarding screens, the single accent across the app, titled error cards, rendered math, hardware example prompts, the stable Settings dialog, and the tidier Preview toolbar.

If you are on beta.17 or beta.18, this update installs itself: InteJer downloads it in the background and offers "Install and restart".

A finished-looking InteJer

This release is about polish. Nothing moved, but almost every screen you touch now looks like it belongs to the same product, and the rough edges testers reported are gone.

One look, everywhere. The steel blue from the InteJer mark is now the single accent across the app: the send button, the sign-in screen, the onboarding tour, badges, selections and focus rings all draw from the same palette. The onboarding tour no longer uses its own green, and the UI font is now Inter on every platform, so Windows, macOS and Linux render the same.

A proper sign-in screen. Sign-in opens on a two-panel screen with the brand on the left and the action on the right. While the browser is open you can see that InteJer is waiting, reopen the page, or cancel. Creating an account, the privacy policy and the terms are one click away.

Errors that tell you what to do. When a request fails, the timeline now shows a titled card instead of a bare line of red text. An expired sign-in offers "Sign in again", a spent allowance offers "Open billing", a busy gateway says so, and the raw server message is always one click away under Details.

Math renders. Dimensions and formulas written as $92.0 \times 50.0\,\text{mm}$ now render properly in the agent's replies, while prices like $39/mo stay as they are.

Easier to scan. Each assistant turn starts with a small InteJer marker, your own messages sit in a distinct bubble, and the composer's example prompts are now hardware prompts: datasheets, boards, firmware, probes.

Settings that stay put. The Settings dialog keeps one height as you switch pages. Under Providers, the models that come with your InteJer plan are listed as "Included" and can no longer be disconnected by accident.

Preview and terminal. The Preview panel's file chips scroll in one row, the viewport theme is a Dark / Blueprint / Light switch with icons, and "Open in" is a labelled group for FreeCAD, Blender and KiCad. An open terminal drawer with no terminals explains how to open one instead of sitting blank.

Smaller fixes. The session sidebar can no longer be dragged to an unreadable sliver, and the app menu is labelled InteJer for screen readers.

If you are on beta.17, this update installs itself: InteJer downloads it in the background and offers "Install and restart". Earlier versions need one last manual download.

The updater's last hidden fault, found and fixed

Beta.15 armed the built-in updater; beta.16 was supposed to prove it. It didn't — every check quietly failed with "No published versions on GitHub" while the button looked like it did nothing. The cause was a single leftover configuration line that made the updater look for a release channel that doesn't exist. It's gone.

This is genuinely the last manual download. From this version on, when a new release is published, InteJer downloads it in the background and shows a toast the moment you open the app: Install & restart or Not yet. No more checking this page.

Two related honesty fixes: a failed update check now says "Check failed — try again" instead of silently doing nothing, and if GitHub refuses to list the release notes (some networks are throttled), the panel says it couldn't load them and offers Retry — rather than pretending there are none.

Everything from beta.16 is in here too

If you're coming from beta.15, this includes the beta.16 changes: the Preview panel with official FreeCAD, KiCad, and Blender icons, Open in Blender (with one-click Blender install), file-type badges — and the three specialist agents now run on Gemini 3.7 Flash, which reads attached images directly.

The Preview panel grew up

The CAD Preview panel now looks the part. The FreeCAD and KiCad tabs carry their official icons, files show what kind of file they are at a glance — CAD sources read stronger than the meshes exported from them, because the source is the editable truth — and a small status chip in the viewport tells you exactly what you're looking at.

Open your designs in Blender

Next to "Open in FreeCAD" there's now Open in Blender. It opens the design's mesh export in a clean Blender scene, framed and ready — and if the design only exists as CAD so far, it tells you to ask the agent for an STL export instead of opening an empty scene. Don't have Blender? The button downloads and installs it for you, official installer, one permission prompt — the same way InteJer already installs FreeCAD and KiCad.

The agents think with Gemini now

The three specialist agents — InteJer Agent, Frontend Developer, and InteJer Designer — now run on Google's Gemini 3.7 Flash. It's faster, it has a million-token context window, and unlike the previous model it can see: attach a screenshot or a photo and the agent reads it directly. Their expert advisors are unchanged and still review every plan.

For everyone updating from beta.15

This is the first version delivered by the built-in updater. If InteJer offered you this update and installed it — that system works now, and every future version arrives the same way.

Two new models in the picker

• Gemini 3.7 Flash (via OpenRouter) — Google's fast multimodal model with a

1M-token context window. It can read images and PDFs you attach. Served through InteJer's gateway like every other model: sign in and it works, no API key of your own.

• The DeepSeek V4 entries now show which dated release you're actually

getting — V4 Pro (0813) and V4 Flash (0731).

The updates panel no longer offers you an ancient version

On beta.14 the changelog panel showed "New version v0.1.0-beta.9 available" — to people already running beta.14. GitHub returns this repository's releases in an order that collapses to alphabetical, where "beta.9" sorts above "beta.14" the way "z" sorts above "a1". The app took the top entry at face value. Releases are now sorted by actually parsing the version numbers.

Auto-update, actually armed this time

Beta.14 shipped the updater feed, but two settings would have kept it silent forever: the updater was told to ignore prerelease versions — and every InteJer beta is one, so it would never have offered anything — and it was willing to install downgrades if the feed ever misbehaved. Both fixed.

If you're on beta.14, this download is the last manual one. From beta.15 onward the app sees new versions itself and installs them in one click.

Your existing conversations answer again

On August 14 at 11:19 UTC, every conversation started before that moment stopped working — a prompt would appear in the timeline and simply never get an answer. New conversations were fine, which made it look random. It wasn't: message ids carry a 48-bit internal timestamp that silently wrapped around at that instant, worldwide, so every new message compared as older than every old one. The app concluded the prompt you just sent had already been answered and never called the model.

Fixed by comparing real creation times, which don't wrap. All of your pre-August-14 conversations pick up exactly where they left off.

InteJer can now update itself

Since the very first beta, the built-in updater has been pointed at nothing — every release required finding and downloading the installer by hand, and most people understandably didn't. From this version on, the app checks our releases feed, tells you when something newer exists, and installs it. This is the last version you will ever need to download manually.

See what changed, inside the app

A new Updates button opens a changelog panel with the release history, and after an update installs you get a short "what's new" popup — so a fix like the one above reaches you as more than a version number.

A prompt optimizer in the composer

A new button rewrites a rough prompt into a precise one before it is sent, and when your request is ambiguous it asks you short clarifying questions first — answer them and it folds the answers in. Apply the result or keep refining; nothing is sent to the agent until you say so.

Archived conversations can be browsed and restored

Archiving is no longer one-way: a new dialog lists everything you've archived and puts any of it back with one click.

Smaller fixes

• If syntax highlighting fails for a code block, only that block falls back to

plain text — previously the whole message lost its formatting.

• The copy button on code blocks no longer multiplies as a message re-renders.

File contents are visible again

Code stopped showing up. A tool call in chat would render its header — create_pcb.py +2 −3 — and then nothing underneath. Opening a file beside the Preview panel gave a blank pane. The file was fine and the edit had been made; you just could not see any of it.

The syntax highlighter compiles a small WebAssembly module, and a security policy added back in 0.1.0-beta.8 was refusing to let it. Everything that depends on highlighting — diffs, the file viewer, code blocks — therefore painted nothing at all. It only ever happened in the installed app, never in development, which is why it went unnoticed for five releases.

Fixed by permitting WebAssembly specifically. Nothing else was loosened.

The advisor stops going "unavailable" mid-session

Each agent consults a stronger model — Claude Sonnet 5 or Kimi K3 — for planning and review. After a handful of turns it would start answering "Advisor is unavailable" and never recover, even though the advisor itself was up.

There is a per-session limit on how many consults an agent may make, and failed attempts were being charged against it. So a brief upstream hiccup could quietly spend the whole allowance, and the advisor stayed switched off for the rest of that session. Only successful consults are counted now, and the allowance itself is larger. If you do reach it, the message says so plainly instead of implying an outage.

Long sessions cost less and stay quicker

A session's history was only ever summarised when it neared the model's absolute limit — around a million tokens. In practice that never happened, so a long day's work kept re-sending an ever-growing transcript with every single message.

History is now condensed once a session gets genuinely large, at a natural pause between tasks rather than in the middle of one. Long sessions get noticeably cheaper to keep going, and the agent's memory of what it is doing is unaffected.

Upgrading

Install over your current version — your projects, settings and sign-in are kept.

Known issue

Requests can occasionally be blocked by the network layer in front of our API, surfacing as "Forbidden: request was blocked by a gateway or proxy." It is not your account and not your credentials — signing out will not help. It is being raised with the hosting provider; starting a fresh session is the current workaround.

The agents can see images now

Ask InteJer about a screenshot, a UI mock-up, a PCB or 3D render, or a photo, and it will actually look at it.

Until now it could not. The model that runs your session is text-only, so an image was invisible to it — and rather than say so, it would often reason from the filename and sound confident about a picture it had never seen. The advisor model behind each agent is able to see (Kimi K3 and Claude Sonnet 5 are both multimodal); it was simply never handed the image.

Now the agent passes the file to its advisor and works from what the advisor actually reports: layout and spacing, colours, labels, what is wrong and where. This applies to all three agents — InteJer Agent, Frontend Developer, Designer.

Practical notes:

• Point at the file however you like ("look at screenshots/home.png"), or just

ask about a render the agent produced itself.

• Up to 4 images per consult, 5 MB each, PNG/JPG/GIF/WEBP. Anything skipped is

reported back rather than silently dropped, so the agent can never quietly reason about an image that never arrived.

• Images are sent only when the question is genuinely visual. They cost real

tokens and, unlike the conversation, are not cached.

• Expect a visual consult to take longer than a text one — the vision models are

slower, and a full-screen screenshot is a lot to read.

The composer's attach button is still hidden on Z.AI models; that is separate and unchanged. Point the agent at a file path instead.

Also in this release

Everything from 0.1.0-beta.11, if you skipped it — account sign-in replacing access keys, and the 20 MB attachment cap that stops large files from bloating the app until it crawls.

Known issue

Requests can occasionally be blocked by the network layer in front of our API, surfacing as "Forbidden: request was blocked by a gateway or proxy." It is not your account and not your credentials — signing out will not help. It is being raised with the hosting provider; starting a fresh session is the current workaround.

Action required — access keys no longer work

Sign in with your InteJer account. The access-key box is gone, and any key you were using is now rejected by the server — including on older builds, so updating is the only way back in.

Click Sign in with browser on the sign-in screen. Your browser opens, you authenticate, and the app continues automatically. Every existing tester account is already provisioned; there is nothing to request and nothing to paste.

Why the change: the five shared keys carried no identity. Usage could not be attributed to anyone, per-account limits could not be applied at all, and a key that leaked could not be revoked without cutting off every tester at once. Signing in fixes all three, and it is what makes your usage and settings appear on your dashboard.

Fixed

Large attachments no longer bloat the app until it crawls. Dropping or pasting a file read it into the conversation with no size limit whatsoever, so a single 19 MB PDF became roughly 25 MB of encoded text held in the message, in the page, in storage, and loaded back into memory every time that session was reopened. A few of those was enough to make clicking and typing take seconds.

Attachments are now capped at 20 MB, and refused before the file is read rather than after. This is not a new restriction: the model request already rejected media above that size — it just did so at the very end, once all the cost had been paid. The message now names the file and its size.

Note that a file under the limit is still embedded the same way as before. The cap stops the pathological case, not the underlying design.

Already live — no update needed

These are server-side and already active for everyone:

• Every tester account is on the Team plan. Unmetered GLM; only the InteJer

Agent advisor allowance is counted.

• Gateway capacity. Stopping a generation mid-stream used to leak a

concurrency slot permanently — about ten stops could saturate an upstream, after which every request returned "every upstream is saturated" with no way to recover on its own.

• Plan detection. A brief database read failure silently downgraded a paying

account to Free and refused it outright.

Known issue

Requests can occasionally be blocked by the network layer in front of our API, surfacing as "Forbidden: request was blocked by a gateway or proxy." It is not your account and not your credentials — signing out will not help. It is being raised with the hosting provider; starting a fresh session is the current workaround.

Accounts and billing

The app could show one account while billing another. After signing in, the sidecar kept serving the previous account's credentials. A Team or Pro account was therefore metered — and refused — as though it were on the Free plan, with an upgrade prompt for the plan it already owned. Restarting the app was the only workaround.

Two independent faults, both required to produce it, both fixed:

• The sign-in write was dropped whenever it arrived before the sidecar finished

connecting — which is the normal case at startup — and a dedupe guard then turned every later retry of that same token into a no-op.

• The credential was resolved once and cached for the sidecar's entire life, so

even a successful write could not reach a running server.

Long sessions no longer expire after about an hour. Same root cause: the access token rotates hourly and the running sidecar never picked up the refresh.

Signing out no longer locks you out. The account menu turned into a help icon after sign-out, leaving no way to sign back in or switch accounts.

New

Preview button in the titlebar. The CAD/EDA preview now has a permanent toggle beside the folder icon. It was previously reachable only by opening the side panel and finding the tab.

InteJer's own colour palette. Earlier builds shipped upstream OpenCode's theme with only the name changed. There is now a navy surface ramp and a steel-blue accent derived from the logo, tuned separately for light and dark and checked against WCAG contrast minimums on every surface.

CAD preview

Preview failures now explain themselves. A file containing no solid geometry says exactly that and names the mesh file beside it that does have geometry, instead of surfacing a raw internal error naming an IPC channel you cannot act on. If a preview fails, check the input file before suspecting the viewer.

Already live — no update required

These shipped server-side and are active for everyone regardless of app version:

• Gateway capacity. Stopping a generation mid-stream leaked a concurrency

slot permanently. Around ten stops could saturate an upstream, after which every request returned "every upstream is saturated. Retry shortly." and the client retried against a state that would never clear on its own.

• Plan detection. A transient database read failure silently downgraded a

paying account to Free, producing an immediate hard refusal.

InteJer 0.1.0-beta.9

Two new specialist agents in the picker

The agent picker now offers three agents, each backed by its own expert advisor that sketches a short plan before it starts working:

• Frontend Developer — UI architecture, state management, styling, and rendering performance.
• InteJer Designer — parametric FreeCAD modeling, and 3D-printable part and enclosure design.
• InteJer Agent — full-stack IoT builds, from PCB to firmware to cloud (unchanged).

Whatever agent you're in will point you to the right one for the task in a single line, so you're never stuck in a general-purpose agent for specialist work. Each agent also opens every request by consulting its advisor for a quick plan, so you get a considered approach up front instead of a cold start.

The InteJer theme is now the default

The app opens in the InteJer theme — light and dark, across the window and the titlebar. You can still change themes from settings at any time.

Reliability

• The 3D / FreeCAD preview is steadier when switching between designs: it no longer briefly tries to render a model with the wrong loader.

If you installed beta.7, update. It shipped with a bug that stopped your sessions from loading. This release fixes it.

Your sessions load again

beta.7 could fail on startup with "Failed to load sessions" and show you an empty history. Nothing was lost — the sessions were on disk the whole time; a database migration tripped over a column that already existed and took the whole session list down with it. The migration is now tolerant of work that's already been applied, so it steps over it and carries on.

FreeCAD preview works on FreeCAD 1.0

Previewing a .FCStd file silently produced nothing on FreeCAD 1.0.x. Two separate quirks in that version were swallowing the conversion — it discarded the arguments we passed it, and then discarded the error message that would have told us so. Both are worked around; a real model now renders in the Preview tab.

While fixing this: if a STEP file contains no actual geometry (an empty export — it happens, and the file still looks valid), the Preview tab now tells you "no tessellatable geometry found" instead of failing quietly. If you see that, the file needs re-exporting from whatever produced it.

Stored approvals no longer take the app down

A schema mismatch in the shared database could kill the permission route outright. It now degrades to "no stored approvals" and logs a warning, rather than failing the request.

Also in this build

Everything from beta.7: InteJer Agent pinned to GLM 5.2 (it was running on whatever model you had selected), agents hidden from the model picker, classic layout by default, and the onboarding tour. Account sign-in was fixed server-side and is already live for everyone regardless of build.

Known issue

This build reports 0.1.0-beta.7 in its version string — the version is stamped at package time and lagged by one release. Cosmetic; the binary here is the beta.8 code. Fixed in the next build.

Installing

Windows x64, unsigned beta build — SmartScreen will warn on first launch. Choose More info → Run anyway. Run the installer over your existing install; your settings and sessions are preserved. In-app auto-update isn't enabled during the beta, so pick up new builds here or from the Downloads page in your dashboard.

InteJer Agent now actually runs on the model it's supposed to

Last release introduced InteJer Agent — a frontier model that plans against your full session transcript while a 1M-context model does the work. It had a bug: the agent ran on whatever model you happened to have selected in the picker, not the one it was designed around. If you'd switched models mid-session, you weren't getting the agent as intended.

InteJer Agent is now pinned to GLM 5.2 as its executor, regardless of what's selected. The same goes for the PCB engineer agent, which now inherits a sane model instead of being steered by the picker. Both are hidden from the model dropdown, because picking a model for them was never meaningful — that choice is what broke them.

The picker still controls Build and Plan, which is where it belongs.

Signing in works again

Account sign-in had been failing with a 401 on every model — the gateway was authenticating against a Supabase API key the project no longer honors. This was a server-side fix and is already live for everyone: you don't need this build to get it. License-key users were never affected, which is why it went unnoticed for as long as it did.

Also in this build

• Classic layout is the default. No more hunting for it in settings on a fresh install.
• The debug overlay and the floating "?" button are gone. They were development scaffolding that shouldn't have shipped.
• All five MCP servers (KiCad, PlatformIO, MongoDB, Playwright, Stitch), the PCB Pack skills, and the FreeCAD resources are bundled — no separate setup.

Installing

Windows x64, unsigned beta build — SmartScreen will warn on first launch. Choose More info → Run anyway. Run the installer over your existing install; your settings are preserved. In-app auto-update isn't enabled during the beta, so grab new builds from here or from the Downloads page in your dashboard.

InteJer Agent

A new agent in the picker. Press Tab and choose intejer-agent.

Most coding agents hand your whole task to a single model: the same context window that plans the work also executes it. On a job that runs from schematic to firmware to broker to dashboard, that model runs out of room and starts guessing. InteJer Agent splits the two jobs.

• A frontier model does the thinking. Claude Sonnet 5 reads the full session transcript — everything the agent has seen and done, not a summary — and weighs in at three moments: before work starts (plan), when the agent gets stuck (stuck), and before it calls the job done (verify).
• A 1M-token model does the work. GLM 5.2 does the reading, editing and running, so your KiCad schematic, the firmware that drives it, the MQTT topic map and the dashboard that renders the telemetry all stay in context at the same time.

You get frontier-model judgment on the decisions that actually matter, and a long-context model everywhere else. That is the whole point: a single general-purpose model has to choose between thinking hard and holding your entire stack in its head. InteJer Agent doesn't.

Details

• The advisor lives server-side. Prompts and model routing sit in the InteJer gateway, so the agent gets better without you reinstalling anything.
• It never blocks your run. Offline, signed out, or over your allowance — the executor just proceeds on its own judgment. A consult that fails is a consult you don't get, not a run that dies.
• Consults are metered as the `InteJer Agent` SKU and appear on your billing page. Team plans include 150 runs per seat per month.
• Scoped to `intejer-agent` only. The build, plan and pcb-engineer agents are unchanged and never see the advisor.

Install

Windows 10/11 (64-bit). This build is unsigned — SmartScreen will warn on first launch. Choose More info → Run anyway.

Windows only; macOS/Linux to follow.

• Stability and reliability improvements.
• Includes all fixes from prior betas (PCB design tooling, blank-screen-on-startup, and earlier).

Windows only; macOS/Linux to follow.

• Improved the built-in PCB design tooling (updated, vendored KiCad integration).
• Stability and reliability improvements.
• Includes all fixes from beta.3 (blank/white-screen-on-startup fix, GLM-5.2 in the model picker).

Adds GLM-5.2 to the model picker (GLM-5.1 · GLM-4.7-Flash · GLM-5.2; the older glm-5v-turbo is removed). GLM-5.2 is a distinct text model with a 1M-token context and is routed through untouched by the gateway (no rewrite to GLM-5.1 / 200K cap).

Also includes the blank / white-screen-on-startup fix from beta.2, so this is the recommended build for all users. Windows only; macOS/Linux to follow.

Fixes a blank / white-screen on startup that affected 0.1.0-beta.1 on Windows.

The renderer's IPC bridge (window.api) could fail to initialize under the sandboxed preload, so the app launched but showed an empty window. This build exposes the bridge first and unconditionally, so the UI loads normally — and telemetry can no longer block app startup.

Same application version as beta.1: this is a delivery/stability re-release published under a fresh download URL so every user is guaranteed to receive the corrected binary. Windows only; macOS/Linux to follow.