Skip to content
| Marketplace
Sign in
Visual Studio Code>Programming Languages>PhewNew to Visual Studio Code? Get it now.
Phew

Phew

Preview

Aleksei Aleinikov Labs

| (0) | Free
Bug fixed. Phew. Open-source AI coding agent for VS Code: one-click fixes, review in the editor, checkpoints, chat history, worktree agents, and any model.
Installation
Launch VS Code Quick Open (Ctrl+P), paste the following command, and press enter.
Copied to clipboard
More Info

The Phew mascot looks at you, thinks, reads a file, waits for your answer with a drop of sweat, and wipes the drop away when the bug is fixed

Phew

Bug fixed. Phew.
An open-source AI coding agent for VS Code that fixes errors in one click, shows every change, and works with any model.

License: Apache 2.0 Status: preview

Phew fixed a login bug: the tests pass, and a summary card shows two changed files with Keep all and Undo all

Install from the Visual Studio Marketplace or search for Phew in the Extensions view.

[!NOTE] Phew is in preview. Each release also has the .vsix file on GitHub Releases.

Why Phew

A command runs in a terminal inside the chat: npm test shows passing and failing tests in color, with Stop, Show terminal, and Copy

The welcome screen offers real tasks: continue yesterday's work, fix errors, fix a failed command, review changes

  • Watch every command. Each command of the agent runs in a real terminal inside the chat, live and in color, and the terminal of the chat in the panel shows the same output. Type an answer or press Ctrl+C in either one. Commands of a chat work like one shell: cd and exported variables carry over.
  • Phew it! One click on an error in the editor, a failed terminal command, or failing tests. Phew finds the cause, fixes it, and checks the result.
  • Fixes until green. Phew runs your check command, for example npm test, and keeps fixing until it passes.
  • Review in the editor. Each change has Keep and Undo buttons in the file, and Keep All and Undo All in the editor title. Prefer to approve first? Switch Edits to Ask first in the chat.
  • Undo anything. Each request is a checkpoint. Restore checkpoint reverts all file changes made after it, even after a restart.
  • Chat history. Chats are saved per workspace, named automatically after the first answer, and grouped by day with a line from each conversation. Pin, search, rename, open them in a new tab, and export them as Markdown. Delete has Undo. Every message shows its time, and every answer has a Copy button.
  • A mascot that shows the work. The Phew drop blinks, looks up while it thinks, reads along a line while it reads files, and sweats while it waits for your answer. The Phew icon shows a number when a chat needs you. Choose Lively, Calm, or Still in Appearance.
  • Parallel chats. Open a second chat in an editor tab (⧉) and let two agents work at the same time. Phew warns before two chats change the same file.
  • Steer while it works. Keep typing while the agent works: messages wait in a queue and reach the agent at its next step. /btw asks a quick side question without stopping the task.
  • See the thinking and the context. Phew shows the reasoning of models that stream it, and a context meter. Long chats are compacted automatically, or with /compact.
  • Questions and a checklist. The agent asks you with clickable answers when something is unclear, and keeps a checklist of the task. /goal sets the goal of a chat.
  • Clickable files. src/app.ts:42 in an answer opens the file at that line.
  • Survives rate limits. Phew retries rate limits and server errors as long as the provider asks, and explains a missing quota or credit in plain words.
  • Rich context. The paperclip adds the current file, the selection, the problems, the Git changes, or the last terminal output in one click, and shows what exactly it adds. Type @ to attach any file. Paste or drop screenshots. Use Add to Chat in the editor, the explorer, and the terminal.
  • Parallel agents. New Agent in Worktree works on a new branch in a separate folder while you keep coding.
  • Commits your way. Generate commit messages in the Source Control view. See the author and signing that commits use, and set Conventional Commits, Signed-off-by, GPG or SSH signing, a different author, and trailers. Phew never adds co-authors by itself.
  • Code index and memory. Phew finds relevant code by meaning and remembers durable project facts in .phew/memory.md.
  • Agent Skills. Phew uses SKILL.md skills from .phew/skills, .agents/skills, .claude/skills, and .opencode/skills, in the project and in your home folder, the same format as Claude Code.
  • Several folders. In a multi-root workspace, the agent works across all folders.
  • MCP done right. Phew reads .vscode/mcp.json, .mcp.json (Claude Code), and its own settings, supports local and remote (HTTP) servers with OAuth sign-in, and lets you choose the tools of each server and read the logs (Phew: Manage MCP Servers…).
  • Autocomplete and Next Edit. Inline completions and a predicted next edit, with any model.
  • Modes. Ask, Plan, Code, Debug, Review, and your own modes, each with its color. Shift+Tab switches the mode.
  • Any model, one click away. The model menu opens above the input and shows the provider and the context size of each model.
  • You decide how far it goes. Careful, Balanced, or Autopilot in one click, or set edits, commands, and MCP tools one by one. The panel shows the commands of your allowlist and what differs from your defaults.
  • Safe by default. Phew does not read .env files and keys, removes secrets from tool output, and always asks before dangerous commands such as git push --force, terraform apply, or kubectl delete.
  • Any model. GitHub Copilot without a key or with a GitHub token, OpenAI, Anthropic, Google Gemini, Azure OpenAI, Mistral, xAI, DeepSeek, Groq, Cerebras, Together AI, Fireworks AI, OpenRouter, LiteLLM, and local models with Ollama, LM Studio, or vLLM.
  • Know the cost. Tokens and cost per task, with optional budget limits.
  • Open source. Apache 2.0, no telemetry.

The mode menu above the input: Ask, Plan, Code, Debug, and Review, each with a color, a description, and read-only marks

The auto-approve panel: Careful, Balanced, and Autopilot levels, and rows for edits, commands with the allowlist, and MCP tools

Continue anywhere

  • Handoff notes. After each task that changes files, Phew writes a short note in .phew/handoff/: the goal, decisions and why, dead ends not to repeat, next steps, and how to verify. Changed files, commands, and check results come from a journal, not from the model.
  • Any agent can continue. With your consent, Phew points AGENTS.md (and CLAUDE.md and GEMINI.md) to the latest note, so Codex, Copilot, Cursor, Claude Code, and others pick up where Phew stopped.
  • Knows what changed. /continue compares the note with the code: new commits, another branch, or files that someone changed after the note. The agent rereads those first.
  • Open format. Plain Markdown with a small header. See docs/HANDOFF.md.

A new chat continues from a handoff note and notices that a file changed after the note

Make it yours

  • Your assistant. Give it a name, a look, and a character: Partner, Mentor, Quiet, Strict reviewer, or your own words. Tell it how to call you and which language to answer in. The dice picks a look and a name for you.
  • The Phew family. The drop, Cat, Bot, Ghost, Sprout, Nerd, Night owl, Angel, and Ninja. Each one is the same living mascot: it blinks, follows your pointer, sweats while it works, and sighs with relief when your bug is fixed. The send button wears its symbol.
  • Your chat. Bubbles, Classic, or Compact style, your accent color, font, and text size: click Aa in the chat.
  • Phew Wrapped. A weekly card with bugs fixed and time saved, ready to share, and achievements. All stats stay on your computer.

The assistant card: a cat named Mochi with the Mentor character, and the Phew family of looks: drop, cat, bot, ghost, sprout, nerd, night owl, angel, ninja

A Phew Wrapped card: 12 bugs fixed, about 6 hours saved, 18 tasks, a 5-day streak

Get started

  1. Open the Phew view in the activity bar, or press Ctrl+Alt+P (Cmd+Alt+P on macOS).
  2. Choose Connect a model. Click a provider, paste the key (or use the key from your environment), click Load models, pick a model, and click Connect and test.
  3. Ask a question or describe a change. Type @ to attach context.
  4. Optional: set Checks → Verify command, for example npm test, in Phew: Open Settings.

Phew: Get Started opens a short guide.

Modes

Mode What Phew does
Ask Answers questions. Changes nothing.
Plan Explores the code and writes a plan. Changes nothing.
Code Edits files and runs commands with your approval.
Debug Reproduces the problem, finds the root cause, and fixes it.
Review Reviews the current changes. Changes nothing.
Custom Your own instructions, tools, and model, from the settings or .phew/modes.json.

Ask, Plan, and Review are read-only: the mode switch shows a lock for them.

Keyboard shortcuts

Shortcut Action
Ctrl+Alt+P Open the chat
Ctrl+Alt+Shift+P Open a parallel chat in an editor tab
Ctrl+Alt+L Add the selection or the file to the chat
Ctrl+Alt+. Phew it! Fix the problem at the cursor
Alt+] Suggest the next edit
Ctrl+Alt+Y / Ctrl+Alt+Z Keep or undo all changes in the file
↑ in the input Reuse an earlier request
Shift+Tab in the input Switch the mode
Ctrl+C in a command terminal Stop the command (typing answers its prompt)
/btw, /goal, /compact Side question, goal of the chat, compact the chat

On macOS, use Cmd instead of Ctrl.

Settings

Phew: Open Settings opens a settings page with sections for models, the agent, approvals, checks, budget, commits, worktrees, the code index, autocomplete, memory, modes, and skills, MCP, privacy, and the interface. The approval toggles in the chat change only that chat; save as default makes them the default. Notifications, the sound, the font size of the chat, and the send key (Enter or Ctrl+Enter) are in Agent. phew.commands.terminal chooses whether the terminal of the chat opens in the panel with the first command, only on request, or never. Every setting is also available in the VS Code settings under Phew.

Security and privacy

  • Phew sends your prompts, the files that the agent reads, and tool output to the model provider that you choose. Use a local model if your code must not leave your computer.
  • In Restricted Mode, Phew only answers questions. It does not change files, run commands, start MCP servers, or read project rules, memory, or modes.
  • Model profiles come only from the user settings, so a workspace cannot send your API keys to another server.
  • Every tool call is written to the Phew output channel (Phew: Show Log).
  • Phew collects no telemetry. Statistics stay on your computer.

See SECURITY.md to report a vulnerability.

Development

Requirements: Node.js 22 or newer.

npm ci
npm run check            # type check, lint, and unit tests
npm run build            # build dist/ with esbuild
npm run test:integration # run the extension in VS Code with a fake model server
npm run package          # create phew.vsix
npm run images           # render the README pictures from the real chat (needs Playwright)

Press F5 in VS Code to start an Extension Development Host. The architecture is described in docs/ARCHITECTURE.md.

Releasing

See docs/RELEASING.md.

Contributing

Contributions are welcome. Read CONTRIBUTING.md before you open a pull request.

License

Phew is licensed under the Apache License 2.0. Copyright 2026 Aleksei Aleinikov.

If you redistribute Phew or a work based on it, you must keep the LICENSE and NOTICE files, which credit the author. "Phew" and the Phew logo are trademarks of Aleksei Aleinikov; see TRADEMARKS.md. "Visual Studio Code" is a trademark of Microsoft Corporation.

  • Contact us
  • Jobs
  • Privacy
  • Manage cookies
  • Terms of use
  • Trademarks
  • Your Privacy Choices
  • Consumer Health Privacy
© 2026 Microsoft