Known limits

Catalase 0.0.6-alpha.2 is missing some things a coding tool might have, and some features are present in this release but not yet fully verified. This page lists both.

Why this page exists

Feature lists describe what a tool is meant to do. This page describes what it does not do, and what has not been checked. Each item comes from our 0.0.6-alpha.2 feature map, bug report and release notes.

Two labels appear below. Unverified means the feature ships in this release but has not been fully tested against a live environment. Open means a known defect that has not been fixed in this version.

Platform and install

Limit	What it means	Workaround
No macOS build yet	Catalase 0.0.6-alpha.2 ships for Windows (installer) and Linux (AppImage and .deb). macOS is not available yet; no macOS binary is published.	A macOS packaging script exists in the repository, so you can build from source. The platform has not been tested.
Unsigned installer (alpha only)	The current alpha build is not code-signed, so Windows SmartScreen shows "Windows protected your PC" on first run and Linux may warn about an unverified source. The production release will be code-signed.	Choose More info, then Run anyway. Verify the SHA-256 checksum first. The steps are on the download page.
.deb cannot self-update	The Windows installer and the Linux AppImage can update themselves. A .deb package cannot.	Reinstall a newer .deb to upgrade, or use the AppImage.

Editor and terminal

Limit	What it means	Workaround
Code intelligence is Python-only	Diagnostics and go-to-definition work for Python only, through pyright. Other languages get syntax highlighting and AI suggestions, with no language-server features.	None in this release.
No debugger	There is no debugger in this release.	None in this release.
One terminal per window	Each window has one PowerShell terminal. There is no scrollback sort or filter.	None in the sources.
No move, copy or duplicate in Explorer	Explorer can create, rename and delete files and folders. It cannot move, copy or duplicate them.	None in the sources. Rename, create and delete still work from the right-click menu.

Git

Limit	What it means	Workaround
Push only	There is no pull, fetch, merge, stash or conflict UI. Push is the only remote operation.	None in the sources.
Create PR works with GitHub only	Pull requests on GitLab, Bitbucket or Azure DevOps are not supported. Create PR uses the GitHub CLI, `gh`.	Install `gh`, sign in with `gh auth login`, and make sure the project has a GitHub remote. Catalase checks all three before it pushes, so a PR that cannot be created leaves nothing published.
Create PR not run against a live repository	The checks before pushing were tested. The pull request itself has not been created against a live GitHub repository.	Try Create PR on a test repository before you rely on it.

Models and local inference

Limit	What it means	Workaround
llama.cpp installs, but chat still uses Ollama	You can install llama-cpp-python from Settings. Chat requests still route through Ollama, so llama.cpp is not used for chat yet.	Use Ollama for chat.
llama.cpp needs a Python you provide	llama-cpp-python is installed into a Python interpreter you choose, with the one-click Install button or with `pip install llama-cpp-python`.	Pick the interpreter in Settings → Local inference backend, or click Install llama-cpp-python. The install button works without a prior successful check.
Model discovery is Hugging Face only	The dynamic catalogue discovers GGUF models on Hugging Face. Models hosted on Ollama are not discovered.	Pull a model by name in Settings → Local Models (Ollama).
No cancel for some actions	There is no cancel for an Ollama pull, a CLI install, Test Key or an MCP connect.	None in the sources.
Real model installs and paid-key calls not verified	Installing a real model from the Marketplace, and routed calls made with paid API keys, have not been verified end to end.	None in the sources.
Budget cap not verified with real spend	The cost meter shows spend against a cap, default $5.00. It warns at 80% and alerts at the cap. Behaviour at the cap has not been verified with real paid spend.	Local turns cost $0.0000. Watch the meter when you use a paid model.
Local-only mode (unverified)	Local-only mode substitutes a local model when a cloud model is picked, and reports the switch in the chat. Present in this release but not yet verified end to end.	None in the sources.
Subscription CLIs (unverified)	Subscription plans drive the provider's own CLI. Present in this release but not yet verified end to end.	None in the sources.
Image input not tested with a vision model	You can attach images in regular chat. Vision input has not been tested with a vision-capable model.	None in the sources.
Inline suggestion speed	Suggestions take roughly 50 to 200 ms with a dedicated GPU, and 0.5 to 3 s on CPU only. Speed has not been verified against a range of hardware configurations.	On CPU only, use a smaller model. Catalase recommends one.

Agent and commands

Limit	What it means	Workaround
Approval modes and Undo not verified end to end	The approval modes and Undo on action cards are present in this release. Approvals have not been verified against a full range of scenarios.	Start a session branch in Source Control before a larger task, so its changes are easy to review or discard.
`/diff` does nothing (open)	`/help` lists `/diff`, but it opens nothing and prints no message. Bug B12.	Click the file in the Source Control panel to open its diff.
`/config`, `/settings`, `/resume`, `/continue` toggle	These commands toggle their panel. The first run opens it, and running the same command again closes it.	Run the command a second time to close the panel.
MCP has no management screen	MCP servers are set up by file, in `.catalase/mcp_servers.json`. There is no screen to manage their tools.	Run `/mcp` to check that servers connected.
Two same-named MCP tools break	If one server offers two tools with the same name, one of them becomes uncallable.	None in the sources.
AGENTS.md read-back not confirmed	It is not confirmed that `AGENTS.md` is loaded as context after `/init` writes it.	None in the sources. Do not rely on it until it is confirmed.
`/remember`, custom commands and retry not verified end to end	Project memory, your own commands in `.catalase/commands`, and retry with the next model are present in this release but not yet verified end to end.	None in the sources.
No multi-agent work	Concurrent multi-agent execution is not present in 0.0.6-alpha.2.	None.
No git worktrees per task	A session gets its own branch, in the same checkout.	Use a session branch.
No skills, hooks or plugin system	There are no skills, hooks or lifecycle events. The Marketplace installs models, not extensions.	None in this release.
No headless mode or SDK	There is no scripting, RPC or SDK interface, so Catalase cannot run in CI.	None in this release.
No web or mobile client	Catalase is a desktop app. There is no web or mobile client, and no remote access.	None in this release.
No web search tool	The agent has no web search or fetch tool. `/deep-research` is a prompt template, not a search tool.	None in this release.

Memory and conversations

Limit	What it means	Workaround
Rewind and Fork work on chat turns only	Rewind reverts your files and the conversation, and asks first. Fork starts a new branch from a message and leaves your files as they are. Neither works on Agent-mode turns.	For Agent turns, use Undo on the action's card.
No permanent delete for conversations	A conversation can be archived, and Show archived lists archived ones. It cannot be permanently deleted.	Archive it.
Starting a second conversation	The release notes list "starting a second conversation requires switching to a new one first" as open. The notes do not describe the cause.	None in the sources.
No context compaction	A long conversation is not compacted. The context meter shows how much of the model's window is in use.	Start a new conversation with `/clear` when the meter fills.
No session sharing or export	There are no share links, HTML export or gist export. Memory export is the only export.	None in the sources.
Snapshots are experimental, and quitting waits	Project snapshots are experimental. Every quit waits up to 6 seconds for the closing snapshot, and a snapshot takes about 4 seconds to render. Earlier versions did not wait.	Let the window close on its own. The user guide asks you not to end the process by hand.

Reliability

Limit	What it means	Workaround
No automatic restart after a backend crash	If the backend stops, the status line reads Backend's gone quiet — try restarting Catalase. Catalase does not restart the backend itself.	Restart Catalase. Your files and conversations are kept.
Status can read Connected after a dropped stream	After a mid-stream connection drop, the status line can still read Connected.	None in the sources.

See the changelog →Download for Windows