How Catalase compares

The short answer

If you want an editor rather than an agent, open-source client code, local models as the default path, and no sign-up, Catalase is the closest match in this group. Zed is the other open-source desktop editor here and is further along in maturity and platform coverage. Cline and Aider are strong local-model tools but are agents that attach to an editor you already have, not editors. GitHub Copilot is the most polished and the most locked-in: proprietary, hosted models only, and a GitHub account is required. Two tools people still find in older comparisons — Void and Continue — are no longer actively maintained.

Catalase's honest disadvantages: it is an alpha, there is no macOS build, code intelligence is Python-only, git is push-only, and the installer is unsigned.

What each tool actually is

Form factor matters more than feature counts here. An editor replaces your editor; an extension or CLI adds to it.

Platforms and local models

If you want to run models on your own hardware, or you are on Linux, this is the table that matters.

What only Catalase documents

These features appear in Catalase's own documentation but are not documented in any of the eight tools' vendor pages checked for this comparison. "Not documented" means absent from the vendor's own pages, not that the feature does not exist — it may be in progress, unreleased, or simply not written up. Verified against vendor documentation, November 2026, not independently tested.

• Agent Office. A persistent panel showing what the agent is doing in real time — a visual room-based view (work, break, game, washroom) with a status line and an Open Chat button to jump straight back to the conversation. No equivalent is documented in Cursor, Copilot, Zed, Continue or Cline's own pages.
• Spend cap that switches to local. Catalase lets you set a session cost ceiling. When a turn would breach it, the request goes to a local model instead of the cloud provider — it does not stop, and it does not reject the turn. Cursor's additional-usage limit stops charges; Copilot's budget is a hard stop; Cline's SDK stops on limit. None switch automatically to local.
• Named accounts per provider, isolated. You can register multiple API credentials for a single provider — for example a "Work" key and a "Personal" key under Anthropic — and give each a name. Each account gets its own isolated credential store; they appear as distinct entries in the model picker. Zed supports multiple OpenAI-compatible endpoints and Cline supports separate config directories, but neither documents named isolation within the same provider.
• Forced local-only mode. A single toggle in Settings routes every request to local models and disables all cloud network calls at the application level, not as a provider setting. Cursor's Privacy Mode controls training data use, not network calls. Zed has an AI off-switch. Neither is a forced-local routing control.

What several tools also do

These features are present in Catalase and are also documented in at least one other tool in this group. This is not a disadvantage — it means you lose nothing if you switch — but it is not a reason to choose Catalase over those tools specifically.

• Conversation rewind (checkpoints). Cursor checkpoints files only; Copilot checkpoints files and history together; Cline checkpoints files, the task, or both. Catalase rewinds the conversation. Zed removed its rewind in version 1.27.0; Continue does not document one.
• Fork a conversation. Copilot supports fork; Continue supports /fork from the current history. Catalase also forks.
• Built-in browser. Cursor and Copilot document a built-in browser for agents. Catalase also embeds one.
• Hardware-fit model labels. Cline Desktop labels local models with hardware-fit badges from a fixed list. Catalase rates its curated rows against your actual RAM and VRAM at open time; Hugging Face rows show "Unknown fit" because the metadata needed for the check is not available there. This is a difference in depth, not an exclusive feature.
• Different model per task. Zed, Copilot (partially), Continue (via model roles) and Cline (Plan vs Act modes) all support routing different tasks to different models. Catalase supports five task types: Chat, Commit messages, PR title/body/changelog, Explain diff and Code review.
• Drive an existing subscription CLI. Zed supports external agents via ACP; Cline supports a Claude Code provider; Copilot has Claude and Codex harnesses. Catalase drives seven coding-plan CLIs — Claude Pro/Max, GitHub Copilot, ChatGPT Plus/Pro, Google AI Pro/Ultra, Alibaba Coding Plan, Amazon Q Developer and opencode.
• No-account local use. Zed's Personal plan is documented as $0 forever with no card, and Cline can run against a local Ollama instance. Catalase has no login in the product at all.

Where Catalase is behind

These are the gaps in v0.0.6-alpha.2.

• It is an alpha. Version 0.0.6-alpha.2. Cursor, Copilot and Zed are mature products with far more usage behind them.
• No macOS build yet. Windows and Linux only. Zed, Cursor, Cline and opencode all ship macOS builds.
• Code intelligence is Python-only. Diagnostics and go-to-definition work for Python via pyright. Other languages get syntax highlighting and AI suggestions, but no language server. Cursor and Zed are broad here.
• Git is push-only. No pull, fetch, merge or stash. Create PR requires the GitHub gh CLI and works with GitHub only.
• The installer is unsigned. Windows SmartScreen warns on first run; Linux may warn about an unverified source.
• No headless mode or SDK. If you want to script an agent in CI, Aider, Cline, opencode and the Copilot CLI all do that and Catalase does not.
• One terminal per window, and no debugger.
• Local inference needs Ollama installed separately.

Where Catalase is ahead

• A real editor with open-source client code and no account. In this group only Zed matches that combination, and Zed is copyleft rather than permissive.
• Local models as the default path, not an add-on. Inline completion has no cloud network client at all — structurally, not as a setting.
• Spend controls that act. At the cap, Catalase switches to a local model instead of stopping or continuing to spend. No other tool in this group documents the same behaviour.
• Hardware-fit model selection against your current RAM and VRAM. The Marketplace rates models against the memory actually free on your machine at open time. Cline Desktop also shows fit labels, but from a fixed list rather than a live hardware check.
• It can drive the subscription you already pay for — seven coding-plan CLIs — without ever seeing your credentials, because you sign in to the vendor's own CLI. Zed, Cline and Copilot document similar capabilities for one or two CLIs each.
• Named accounts per provider. Register a Work key and a Personal key under the same provider; they stay isolated and appear as distinct entries in the picker. No other tool in this group documents the same at the provider level.
• Permissive licence with an explicit fork policy. Apache-2.0, with the name and logo excluded so a rename is clean.

Sources

Each claim above comes from the vendor's own material, checked in November 2026 and not independently tested. Where a vendor's page was silent, the cell is marked for confirmation rather than filled in. The unique-features and also-available sections are drawn from these same pages.

• Cursor — cursor.com/docs/agent/chat/checkpoints (checkpoints/rewind); cursor.com/docs/agent/browser (built-in browser); cursor.com/docs/models (model list); cursor.com/docs/enterprise/privacy-and-data-governance (Privacy Mode); changelog. Repository is not public; pricing could not be attributed to the editor at the time of checking, as noted above.
• GitHub Copilot — plans page (Free $0 with 2,000 completions per month and no credit card; Pro $10; Pro+ $39; Max $100); docs.github.com/en/copilot/concepts/billing/budgets-for-usage-based-billing (budget/hard stop); code.visualstudio.com/docs/copilot/chat/chat-checkpoints (chat checkpoints); /docs/agents/run/sessions/manage-sessions (sessions list); /docs/debugtest/integrated-browser (built-in browser); /docs/copilot/customization/language-models (model routing); /docs/agents/run/agent-harnesses (Claude and Codex harnesses).
• Zed — zed.dev/docs/ai/agent-panel; zed.dev/use-a-local-model; zed.dev/docs/configuration; zed.dev/docs/agent-settings; zed.dev/docs/external-agents (ACP agents); zed.dev/releases (rewind removed in 1.27.0). Pricing page: Personal $0 forever, Pro $10/month, Business $30/seat. Licence: GPL-3.0-or-later with Apache-2.0 components. Site states macOS, Linux and Windows.
• Continue — docs.continue.dev/cli/tui-mode; /customize/model-roles/00-intro (model roles per task); /customize/models; github.com/continuedev/continue README (/fork command). Repository states it is no longer actively maintained and read-only; Apache-2.0; described as a CLI, VS Code extension and JetBrains plugin.
• Cline — docs.cline.bot/usage/cline-desktop-local-models.md (hardware-fit badges, desktop); /core-workflows/checkpoints.md (file/task/both checkpoint); /core-workflows/plan-and-act.md (Plan vs Act, model routing); /features/auto-approve.md; /getting-started/config.md (separate config dirs); /provider-config/anthropic.md (Claude Code provider). Apache-2.0; SDK, IDE extension, CLI and desktop app for macOS and Windows; provider list includes Ollama and LM Studio.
• Void — repository archived 2 June 2026, read-only; README states deprecation; Apache-2.0.
• Aider — Apache-2.0; site describes terminal pair programming and connecting to "almost any LLM, including local models".
• opencode — MIT; terminal UI and CLI with a beta desktop app for macOS, Windows and Linux.
• Catalase — this repository: apps/desktop/package.json for version and licence, docs/README.md for platform, provider and control behaviour, docs/INSTALLING.md for Linux packaging, TRADEMARKS.md for the fork policy, the open-source strategy document for the no-telemetry and no-account audit, and apps/desktop/renderer/src/modelSelector.js for per-provider named-account behaviour.

If you spot a cell that is wrong or out of date, it is a bug in this page. Contact details will be published alongside the repository.