RELEASE_NOTES.md

ShadowCode 2.2.0

ShadowCode 2.2.0 runs on Macs, both Apple silicon and Intel, and its Linux Flatpak is now a download on every release. It also connects to your own model servers and adds find in conversation, a smarter command palette and text size. The 2.1.0 notes follow below.

New

• ShadowCode for macOS. ShadowCode_2.2.0_aarch64.dmg for Apple silicon and ShadowCode_2.2.0_x64.dmg for Intel Macs (macOS 12 or later). Local models run on llama.cpp with Metal; agent shell commands run in the macOS sandbox, with your home folder and other drives hidden; API keys can be kept in the Keychain. Install or update with one command, or open the .dmg (see the README; a copy downloaded in a browser needs Open Anyway in Privacy & Security the first time).
• The Flatpak on every release. ShadowCode_2.2.0_x86_64.flatpak and ShadowCode_2.2.0_aarch64.flatpak, with the Vulkan GPU runtime, built, installed and started from the release before they are attached.
• Your own model servers. Settings › Accounts › Your own endpoints connects Ollama, LM Studio, llama.cpp or vLLM servers, xAI, OpenAI or any OpenAI-compatible API; the models you pick run with ShadowCode's own agent.
• Settings › Agent. Steps, budgets, time limits, retries, history shortening, checkpoints and more, each with its range, default and Reset.
• Find in conversation (Ctrl+F), search that also matches answers, a command palette that matches letters in order and lists conversations and projects, text size (Ctrl+=, Ctrl+-, Ctrl+0), message times, and Show all and Copy for long tool output.

Fixed

• Approval cards show the folder a command runs in. Long paths such as /Volumes/Disk2/projects/app could be hidden as if they were secrets.
• Low memory is named before a model loads. A local model that needs more memory than is free right now says so, instead of running very slowly.
• The theme follows the desktop by default, and screen readers hear milestones rather than every streamed update.

Download

Linux x86_64: ShadowCode_2.2.0_amd64.AppImage, ShadowCode_2.2.0_amd64.deb and ShadowCode_2.2.0_appimage-runtime-sources.tar.gz, with SHA256SUMS, RELEASE-MANIFEST.json, RELEASE-AUTH and RELEASE-AUTH.sig. For ARM64 the packages end in _aarch64 (AppImage) or _arm64 (.deb), with signature files ending in -aarch64. Install the AppImage with the authenticated installer described in the README; the signature is checked before anything runs. The Flatpak bundles and the macOS disk images are attached shortly after publication, with their SHA-256 checksums in ShadowCode_2.2.0_flatpak.sha256 and ShadowCode_2.2.0_macos.sha256.

ShadowCode 2.1.0

ShadowCode 2.1.0 runs on 64-bit ARM computers, installs and updates from a signed APT repository, and its installer now checks that the new version starts before it keeps it. The 2.0.0 notes follow below.

New

• 64-bit ARM. The AppImage, the Debian package and the Flatpak are built and tested on ARM64 (aarch64) as well as x86_64, and each is signed for its own architecture. Download the files ending in _aarch64 or _arm64.
• Updates through APT. Add the signed ShadowCode repository once and apt upgrade keeps ShadowCode current, on both architectures. The README's Debian section shows how, including how to check the repository's key.
• An installer that checks before it commits. The AppImage installer saves a copy of your database, starts the new version and waits for its window. If it hangs, crashes or reports the wrong version, everything is put back as it was, database included.
• Model history in the picker. Each model shows how long tasks took in this project, what they cost when the provider reported it, and how often their checks passed. Compare history shows them side by side, and an optional suggestion points to the cheapest model that passed here before. ShadowCode never switches models on its own.

Fixed

• The window remembers its size again, including on Wayland desktops such as KDE Plasma, without growing from launch to launch.
• Website allow-list in the Flatpak. Agent shell commands in the Flatpak can now be limited to the websites you allow; if that can't be set up, the command is refused rather than run unrestricted.

Download

For x86_64: ShadowCode_2.1.0_amd64.AppImage, ShadowCode_2.1.0_amd64.deb and ShadowCode_2.1.0_appimage-runtime-sources.tar.gz, with SHA256SUMS, RELEASE-MANIFEST.json, RELEASE-AUTH and RELEASE-AUTH.sig. For ARM64 the packages end in _aarch64 (AppImage) or _arm64 (.deb), with signature files ending in -aarch64. Install the AppImage with the authenticated installer described in the README; the signature is checked before anything runs.

ShadowCode 2.0.0

Version 2.0.0 adds a source-built Flatpak, closes three reproduced remote-access defects, and updates the affected development dependency. See Flatpak for installation, host-tool permissions, and package-specific limits. The source is published in this repository; binary release publication and Flathub submission remain pending. About ShadowCode shows the version, Apache license, publisher, official website, repository, and contributors. The trademark and branding policy covers the official identity and permits forks and accurate community attribution. Previous release notes follow.

ShadowCode 1.0.1

ShadowCode 1.0.1 fixes the size of the window: it no longer grows each time it opens on Wayland, and it fits small screens. The 1.0.0 notes follow below.

Fixed

• The window no longer grows each time it opens. On Wayland desktops such as KDE Plasma, the window opened a little bigger every time (about 50 pixels wider and 100 taller) until it no longer fit on the screen. ShadowCode no longer remembers the window's size on any desktop, so it opens at its default size, fitted to the screen. It still reopens maximized or full screen if you left it that way, and on X11 at the same place. If your window already grew, it opens at its normal size again.
• The window fits small screens. On a 1366x768 screen the window used to be taller than the screen, so the composer and the status bar were hidden under the panel. The window now opens at a size that fits the screen, with room for the title bar and the panel. With more than one screen it fits the screen it reopens on (X11), or every screen (Wayland, where the desktop chooses). Larger screens still get the full size, the window opens maximized only when no screen is big enough for it, and a window you left maximized opens maximized.

Download

ShadowCode_1.0.1_amd64.AppImage, ShadowCode_1.0.1_amd64.deb and ShadowCode_1.0.1_appimage-runtime-sources.tar.gz, with SHA256SUMS, RELEASE-MANIFEST.json, RELEASE-AUTH and RELEASE-AUTH.sig. Install the AppImage with the authenticated installer described in the README; the signature is checked before anything runs.

Published 1.0.0

ShadowCode 1.0.0 made ShadowCode easy to start and ready for real work. New users get free models first, plain explanations when something fails and a glossary of the words the app uses. Everyone gets approvals that explain themselves, secret checks before anything leaves the computer, one rulebook for every agent, second opinions from another model, spending limits, and tools for big projects.

Easy to start

• Free models first. Until you connect a subscription, the model picker lists the free models on this computer first. Your code stays on your machine.
• Plain words when something fails. If a task fails because a key was refused, credits ran out, the conversation got too long, the model doesn't exist, the provider is busy, the local model isn't running or you're offline, ShadowCode says so simply and offers the next step, such as Try again or Continue on another model….
• Words you'll see. Press ? for Help: it explains worktree, checkpoint, rewind, context, tokens and more.
• Approvals you can read. Every approval card says what the action does in one sentence, how much it can affect, and whether Rewind can undo it, even for long shell commands and for Codex, Claude Code and the other subscriptions.

Safe by default

• No secrets by accident. Commits, pushes and pull requests from ShadowCode are checked for keys, passwords and .env files first; you choose to remove the file, ignore it, or go ahead.
• Careful with new packages. When the agent wants to install a package, ShadowCode checks that it exists, how new it is, and whether its name is a near-copy of a popular one.
• Your keys in the keyring. API keys can be moved into the system keyring (GNOME Keyring, KWallet, KeePassXC) with one click.
• Heads-ups. If a task skipped or deleted tests, changed CI or switched off a check, the conversation says so. Only change these keeps a task to the files you name.
• Always allow here for the exact test and build commands you run all day, per project, and easy to take back.

Ready for real work

• One rulebook for every agent. Write your rules and skills once; they reach ShadowCode's own agent and every subscription CLI.
• Second opinions. Have another model review your staged changes or a task before you commit, with each finding on the line it's about.
• Roles. Use one model to plan, another to implement and a third to review.
• Spending limits. Paid models ask before spending more than $1 on a task or $10 in a day (you can change both), and show a price before you send.
• Recovers on its own. Busy providers are retried with a visible count; after a plan limit, a task can resume when the plan resets; a stuck task pauses and asks you.
• Big projects. The code index is kept between runs, covers up to 250,000 files and can focus on one folder of a monorepo.
• Review by risk. Changed files are grouped with config and dependencies first, and Explain this change describes a file's change in plain words.
• Parallel tasks that just work. A task in its own worktree can start from any branch, runs your setup commands, gets your .env and its own port.

Reliable

• Your data is safe. Back up, restore, repair or start over from Settings › Your data. Upgrades from every release since 0.28 are tested.
• A model that runs out of graphics memory offers a smaller context; a task that fails early ends with its reason; subscription usage is kept when a turn fails; Rewind keeps files you saved during a subscription turn.
• A stable, documented API for version 1.

Download

ShadowCode_1.0.0_amd64.AppImage, ShadowCode_1.0.0_amd64.deb and ShadowCode_1.0.0_appimage-runtime-sources.tar.gz, with SHA256SUMS, RELEASE-MANIFEST.json, RELEASE-AUTH and RELEASE-AUTH.sig. Install the AppImage with the authenticated installer described in the README; the signature is checked before anything runs.

Published 0.34.2

ShadowCode 0.34.2 made ShadowCode ready for everyone: you can start with no account at all, the app tells you when an update is out, and a full quality pass fixed dozens of bugs across the engine, the window, the vendors and security.

Start with nothing

• No account needed. Open a project and ShadowCode offers a free model that runs on your own computer, picked for your memory and graphics card, with its size and download time. You can also add an OpenRouter key or sign in to a subscription.
• Five free models to choose from in Settings › Local models, from a 2 GB model for small laptops to a 19 GB one for big GPUs. Each download is checked against a pinned fingerprint, can be paused and resumed, and never starts on its own.

Easier every day

• A calmer composer. Effort, Compare and Worktree now live under More, so the message area stays clear.
• Clearer when something goes wrong. Errors read as plain sentences with a Try again button, and the status bar says "Not connected" when the engine is down.
• Keyboard friendly. Escape closes every dialog, and the Tasks panel can be driven entirely from the keyboard.
• Update notice. ShadowCode checks for a new release at most once a day, sends nothing about you, and tells you how to update your kind of install. Settings › About shows the version and license.

Works with today's tools

Checked live against the current Codex, Claude Code, Cursor, Grok and Antigravity releases. Claude Code now shows its real models and plan usage, effort settings work for Claude Code, Codex and Grok, and OpenRouter shows your account balance.

Safer and more reliable

• Rewind and Undo can no longer delete files that existed before a task.
• Slow local models on a CPU are no longer cut off mid-answer, and a long prompt can take up to 15 minutes to start without being cancelled.
• A security audit closed ways secret files could be read, tightened "Allow for this task", blocked DNS rebinding against remote access, and found no known vulnerabilities in dependencies.
• The Debian package passes lintian cleanly and is ready for distributions, with a man page and shell completions.

Download

ShadowCode_0.34.2_amd64.AppImage, ShadowCode_0.34.2_amd64.deb and ShadowCode_0.34.2_appimage-runtime-sources.tar.gz, with SHA256SUMS, RELEASE-MANIFEST.json, RELEASE-AUTH and RELEASE-AUTH.sig. Install the AppImage with the authenticated installer described in the README; the signature is checked before anything runs.

Published 0.33.1

ShadowCode 0.33.1 was published on September 28, 2026 from e15c4480e65db5650af012bb2a9773dbe89acf84. It improves editing, explicit verification, provider sign-in recovery and long-running desktop reliability. Historical notes below retain their original versions and scope.

Everyday workflow

• Editor. Code highlighting, line numbers, undo/redo and recoverable drafts. Saves check the original disk revision, preserve supported newline formats and retain edits typed while a save is pending.
• Context. Inspect the files and bounded excerpts attached to the next task, including exclusions, truncation and approximate token counts. Estimates are not provider-reported usage.
• Run a check… Completed-task summaries can start a fresh, user-entered command against current project files without selecting a model or spending another model turn. Existing trust and command approvals apply. Output and a new receipt appear in a new task while your composer draft and earlier evidence remain intact. A successful command proves that check's result, not overall application correctness.
• Accounts. Returning to Settings restores each provider's active login independently. Cancellation shows Stopping… until acknowledged. Unknown progress preserves cached information instead of silently starting fresh probes. A successful login process alone does not establish Ready or subscription billing.

Workspace and desktop reliability

• Compare refuses app-owned unsaved editor drafts before taking its snapshot. Short writes coordinate across cooperating ShadowCode processes, and interrupted cleanup retains ownership evidence and recoverable work. Arbitrary terminal commands and external applications remain outside that cooperative lock.
• Shared native event registrations avoid growth on every task. A source-built X11 run passed 100 qualification tasks after 25 warm-up tasks with stable listener, descriptor and thread counts and intact SQLite integrity.
• AppImage launch preserves an explicit display-backend choice. The package normalizer accepts linuxdeploy's single exact matching version field while rejecting unrelated launcher changes.
• Approval restoration, provider event ownership, cancellation, local-model scheduling and installer recovery have additional regression coverage.

Release and qualification status

The runtime builder now preserves tracked source pins and reference metadata during ordinary builds and notices-only refreshes. Generated runtime metadata retains the actual compilation time. Shell-flow regressions cover the two input mutations that stopped the unpublished 0.33.0 release; the immutable-source guard remains enforced.

Release jobs separate build, signing and publication, verify private artifact snapshots and require complete scoped receipts. The AppImage installer requires publisher signatures and durable accepted-state checks; checksum overrides and --unverified do not bypass authentication. Previously published unsigned artifacts are not retroactively authenticated.

Recorded source evidence includes 507 UI unit tests, 75 browser scenarios and the 100-task native endurance run. These counts describe the tested source snapshots, not automatically the final 0.33.1 package. Small live coding tasks have passed across Codex, Claude, Cursor, Grok and Antigravity; full live login-expiry/cancellation/continuation qualification and ACP billing entitlement remain separate.

Checks 36418625715 and Native desktop 36418625663 passed on the release commit. Release run 36427024374, attempt 1, passed all 15 required build gates and protected signing. Its publication job failed when the new draft was briefly absent from the release listing, so the workflow retains a failure conclusion. Publication resumed locally using the same frozen publisher at e15c448 and immutable signed artifact 10987931982; all seven remote asset files were downloaded and compared before the release became public. The tag and package bytes were preserved. A fresh download of all seven public assets matched the signed artifact byte for byte, and offline verification of the public AppImage passed against the independently reviewed trust bundle, expected version and full commit.

Physical COSMIC/Wayland interaction coverage is partial: the overall run failed its requested window-resize assertions. GPU inference endurance and broader keyboard/clipboard/IME/DPI acceptance are not established. See the implementation evidence ledger for exact snapshots, retained failures and later results.

The published Linux release includes ShadowCode_0.33.1_amd64.AppImage, ShadowCode_0.33.1_amd64.deb and ShadowCode_0.33.1_appimage-runtime-sources.tar.gz, together with SHA256SUMS, RELEASE-MANIFEST.json, RELEASE-AUTH and RELEASE-AUTH.sig. Use the release procedure and publisher authentication contract to verify packages with independently trusted installer tooling.

Failed 0.33.0 release attempt (unpublished)

Release run 36414378697, from 9e385cb / v0.33.0, failed the managed-runtime source-integrity gate: the builder changed tracked tools/llama.cpp.pin and packaging/llama.cpp/COMMIT timestamps. The runtime build and six runtime tests passed, but signing and publication were skipped. The tag is retained unchanged; 0.33.1 was built and qualified separately. Local package and latency evidence for 0.33.0 is not relabelled as 0.33.1 evidence.

Published 0.32.0

ShadowCode 0.32.0 is the biggest update yet: it's easier to use every day, the agent is smarter, commands are properly sandboxed, and you can use ShadowCode from your phone, your editor, or on a schedule. The 0.31.1 notes follow below.

Easier every day

• Type @ to attach files and folders, ↑ for earlier prompts, and switch between Code, Plan and Ask right in the composer. Models that support it get an effort control.
• Edit and resend any of your messages, or retry or copy it.
• Approvals show the actual change: the diff for a file, the full command for a shell step. Allow for this task stops repeat prompts, and Deny with note tells the agent why.
• Review just what a task changed, full width, and keep or undo each change. Rewind now asks first, can be undone, and also covers shell commands and Codex, Claude Code, Cursor, Grok and Antigravity edits.
• Run tasks side by side: Worktree next to Send starts a task in its own copy of the project; apply it, keep it as a branch or discard it.
• Notifications when a task needs you, fails, hits a plan limit or finishes; click to jump to it. The sidebar marks what needs attention.
• See context and cost for every model in the status bar.
• Preview the app you're building inside ShadowCode and click an element to show the agent exactly what to change.
• A real terminal, a Git tab that goes from commit to pull request with live CI checks, and voice dictation that runs on your computer.

A smarter agent

• Subagents explore, plan, review or work in parallel copies of your project.
• Language servers tell the agent about errors right after each edit, and a repo map and code search help it find its way.
• Long conversations are summarized instead of cut, OpenRouter requests use prompt caching, and dropped connections are retried.
• It reads CLAUDE.md, Cursor rules and Claude Code skills and commands, and your MCP servers reach every model.

Safer by default

• Shell commands run in a sandbox that hides your home folder: SSH keys, cloud credentials and API keys stay invisible, and only the project is writable. You can require the sandbox and limit the network to a list of hosts.

Beyond the window

• Remote access: scan a QR code to follow and steer ShadowCode from your phone. Off by default; Tailscale recommended.
• Zed and JetBrains: shadowcode acp makes ShadowCode an agent inside your editor.
• Automations: run a prompt on a schedule, turn a GitHub issue into a task and pull request, or run ShadowCode in GitHub Actions.

Changed in 0.31.1

• Apache License 2.0. ShadowCode is now Apache-2.0 licensed instead of MIT. A new NOTICE file credits Shadowfetch as the original creator, and anyone sharing a copy or a modified version must include it. Releases up to 0.31.0 remain available under MIT.

New in 0.31.0

• Allowance. One button in the status bar shows what's left everywhere: each subscription's usage and reset times, your OpenRouter credits, and the local models ready to run. Only figures the services report; nothing is guessed.
• Keep going when a plan runs out. When Codex, Claude Code or another subscription hits its limit, the same conversation continues on a local model, with a note saying so. Prefer to choose? Set When a plan runs out to Ask me. ShadowCode never retries on the same plan or buys more usage.
• Compare. Press Compare next to Send to give one task to 2 or 3 models. Each works in its own copy of your project (uncommitted work included); see their changes, checks, time and cost side by side, then Keep the best one. Only the kept changes reach your files, nothing is committed, and the copies are cleaned up. Wins in this project tracks which model you keep. Tested live: two OpenRouter models fixed the same bug in 11 seconds and the kept fix applied cleanly.

Fixed in 0.30.2

• Web for OpenRouter. The Web toggle now appears for OpenRouter models, as the 0.29 notes promised. Tested live: an OpenRouter model searched the web and cited its source.
• Your OpenRouter key stays put. It is stripped from the environment of every subscription tool ShadowCode starts, like the other provider keys.
• No secrets in the repository. A new check scans every tracked file for real-looking API keys and runs on every push.
• Docs for OpenRouter, Antigravity, data locations and testing are current.

Fixed in 0.30.1

• Antigravity works end to end. Tested live after signing in: it lists its 11 Gemini models, asks ShadowCode before running a command, answers, and resumes the conversation. A timing bug that made the first prompt end with no answer is fixed.
• Answers appear once. Replies from Codex, Claude Code, Cursor, Grok and Antigravity are no longer repeated under Result.
• Clearer notes. The model note reads "Using Cursor · Auto · Cloud", and a question that changed nothing ends with one quiet line instead of an empty report card.

New in 0.30.0

• Antigravity asks first. It now runs through Google's official ACP agent server instead of the agy CLI's print mode, which silently denied any command. Commands and edits reach ShadowCode's approvals, images work, and conversations resume.
• One-time install. Settings › Accounts › Antigravity › Install downloads the server from Google (334 MB, checksum-verified). Connect then signs in with Google in your browser. This sign-in is ShadowCode's own and separate from the agy CLI.
• No surprise browser tabs. Status checks and tasks never open a sign-in page. If the sign-in expires, the row says Sign in and a task stops with that message.

New in 0.29.0

• OpenRouter. Paste an OpenRouter API key in Settings › Accounts and pick any of its text models from the picker's new API keys group. Rows show the price per million tokens and whether the model supports images or tools. Every token is billed to your OpenRouter account, and the group says so. These models run on ShadowCode's own agent loop, so permissions, approvals, checkpoints, web tools, images and review all work. See OpenRouter.
• Web search. When DuckDuckGo answers with a bot check, web_search asks Marginalia Search's public API instead of giving up.
• VPN users. Sites that a VPN's DNS resolves into 192.0.0.0/24 (NordVPN does this for Google) are no longer refused by web_fetch.
• Antigravity. Tasks where Antigravity silently denied a command now fail with an explanation instead of reporting success with no answer.

Fixed in 0.28.1

• Accounts. Each card shows the product name, its version once, and the signed-in email with the plan on one line. Usage windows are listed once, with reset times that stay current. Credits appear only when the vendor reports some.
• Reset times round to whole minutes ("resets in 3h", not "2h 60m").
• Stopped tasks that changed nothing show a single quiet line instead of a red card and without a "Needs attention" note. The partial reply is kept.
• Newly added local models always appear in the picker right away.
• Local model names come from the GGUF file's own name, plus its quantization, for example Qwen3 14B · Q4_K_M.
• Notifications appear at the top centre, clear of the composer and the drawer. "Permissions & network" no longer wraps in Settings, and the drawer close button is larger.
• First run no longer checks local servers or vendor tools that the onboarding screen does not use.

ShadowCode 0.28.0

ShadowCode 0.28.0 is a Linux desktop coding agent with one window and one model picker. It runs your existing subscriptions through the vendors' official command-line tools, or GGUF models on your own hardware with a llama.cpp runtime that ships with the app.

What changed

• One picker lists Subscriptions and On this computer. Every row shows Local or Cloud, whether it is ready (and why not), Vision and Chat only badges, and the usage the vendor reports. Each conversation remembers its model.
• Accounts. Settings › Accounts signs in and out with each vendor's official login and logout commands. ShadowCode never reads credential files, and it removes provider API keys from the environment of every vendor task, login and logout.
• Usage. Codex shows its rate-limit windows per quota pool, with reset times. Other vendors show Usage unavailable with the reason. The last snapshot survives a restart. When a vendor reports its plan limit, the task stops and you pick another model.
• Local models. A pinned llama.cpp (Vulkan and every x86-64 CPU level) is bundled in the AppImage and the deb. Models are read from their GGUF metadata. Context is sized to your VRAM or RAM. Models in an existing Ollama store can be imported by reference, without copying and without the daemon. Vision comes from a paired projector.
• Web tools for local models. web_fetch and web_search block private networks and cloud metadata addresses.
• Network modes. Online, Web tools off, Offline.
• Permission modes. Ask before actions (the default for new installs) and Allow project edits. Settings explains what each vendor enforces.
• Switching providers mid-conversation. A bounded summary of the earlier turns is handed over, and ShadowCode asks your consent before local content goes to a cloud provider.
• Resuming vendor sessions. Each vendor's own session is resumed per conversation.
• Installer. scripts/install-appimage.sh now requires SHA256SUMS and installs the llama.cpp runtime from the AppImage into ~/.local/lib/shadowcode, rolling back if a step fails.
• Removed. The legacy Python harness and its tooling.

Subscription status

Vendor	Runtime	Image input	Approvals shown in ShadowCode	Usage
Codex	codex app-server	Yes	Yes	Rate-limit windows per pool
Claude Code	claude -p stream-json	Yes	Yes	Unavailable
Cursor	cursor-agent acp	When advertised	Yes	Unavailable (plan tier only)
Antigravity	agy --print= stream-json	No	No (its own settings)	Unavailable
Grok	grok agent stdio (ACP)	No	Yes	Unavailable (session tokens only)

Local model support

• Formats. Any GGUF whose architecture the bundled llama.cpp supports. Unsupported ones (for example gptoss at this pin) are listed with the reason.
• Hardware. Vulkan GPUs, or any x86-64 CPU. One model is loaded at a time. If the GPU start fails, ShadowCode retries on the CPU and labels the row.
• Tools and vision. Models whose chat template has no tool support run as Chat only. Models with a paired projector accept images and get view_image.

Known limits

• Claude Code was not verified end to end on the build machine, because it was not signed in there. The stream-json adapter is covered by unit tests with scripted protocol frames, and sign-in state by fake-CLI tests.
• Antigravity exposes no machine-readable usage. It never sends approval requests to ShadowCode: its own settings decide what it may do. Sign-in and sign-out happen only inside agy.
• Cursor reports its plan tier but no remaining allowance.
• Grok's ACP interface accepts no images, and Grok has no read-only mode for Plan/Review.
• web_search uses DuckDuckGo's HTML page, which may block automated requests. The tool then reports that no results were retrieved.
• Ollama stores outside ~/.ollama/models are found only through OLLAMA_MODELS or an Environment=OLLAMA_MODELS line in the Ollama systemd user unit. The system-wide service's store isn't found automatically.
• ShadowCode is not an operating-system sandbox. Shell commands run as your user.

Downloads

• ShadowCode_0.32.0_amd64.AppImage
• ShadowCode_0.32.0_amd64.deb
• ShadowCode_0.32.0_appimage-runtime-sources.tar.gz
• SHA256SUMS

Builds are for x86_64 Linux with glibc 2.39 or newer (Ubuntu 24.04 or later). Check the files with sha256sum --ignore-missing -c SHA256SUMS, then install with scripts/install-appimage.sh (see the README).