ManjuAi for VS Code

Privacy-oriented coding agent in the secondary sidebar. It reads and edits files in your open workspace folder. LLM requests are sent to the Manju hosted API when you chat; file edits and the code index stay on your machine.

Features

Guest trial: 2 chat messages without sign-in on Free (OpenRouter auto router), then sign-in is required.
Signed in: GLM 5.3 Flash default (z-ai/glm-5.3-flash) plus GLM 5.3, DeepSeek V3.2, and Grok 4.1 Fast in the model picker. Usage is metered on Manju servers (see the site for plan limits and optional Pro upgrade).
Agent mode: edits files, runs terminal commands (with approval), semantic search, optional web search/fetch when approved.
Local Python stack: agent under python/ (not a bundled binary). On first use the extension creates a venv in extension global storage and runs pip install once. Requires Python 3.10+ on PATH (manju.pythonPath to override).
Local servers (default): integrated terminal runs the agent on 127.0.0.1:8770 and the index worker on 127.0.0.1:8772 (manju.serverLaunch: terminal | extension).
Semantic index: .manju/index/ in the workspace (tree-sitter chunks, Chroma, FastEmbed). Tools: find_relevant_code, get_file_outline. Rules from .manju/rules.md when present.
Chat history: per workspace in .manju/chat/session.json.
Greenfield websites: in an empty folder, “build a website” can copy templates/next-shadcn/ and run npm install, then the agent edits the app.
Approvals: terminal runs and web actions are confirmed in the sidebar before execution.
Secrets: the agent may write .env.example; you add real API keys locally. Sign-in tokens are stored in VS Code Secret Storage, not in workspace settings.

ManjuAi is not affiliated with Microsoft, GitHub, or OpenAI.

Privacy and data

Stays local	Sent to Manju API (when you chat)
Files the agent reads/writes	Your messages and attached context (open files, search hits, outlines)
`.manju/` index and chat JSON	Agent tool payloads and LLM step content (`/manju-chat/local-agent/llm-step`)
Python venv under extension storage	Web search/fetch only after you approve in chat
HTTP servers on `127.0.0.1` (agent/index)	Sign-in handoff exchange (`POST /auth/desktop/exchange`)

Default API base: https://manjuai.tech/manju-api. Override with manju.backendUrl for local backend development.

Guest send counts may also be stored under ~/.manju/vscode-guest-trial.json.

Requirements

VS Code 1.85+
Python 3.10+ on PATH
Network access to your configured Manju API and (on first run) PyPI for pip install

Use

Open a project folder in VS Code (or Cursor with the same VSIX).
Open the ManjuAi view in the secondary sidebar.
Send a message; local servers start in a terminal when needed.
Sign in from the sidebar for metered GLM / DeepSeek / Grok models.

Package and install (`.vsix`)

cd manju-vscode
npm install
npm run check
npm run package

Produces manju-ai-<version>.vsix in this folder.

Install for testers:

code --install-extension manju-ai-0.1.12.vsix

Or: Extensions → ... → Install from VSIX…

Publish (after creating publisher manjuai on the Marketplace):

npx vsce login manjuai
npx vsce publish

Or upload the .vsix manually at Manage publishers.

After web sign-in, redirect to:

VS Code:

vscode://manjuai.manju-ai/auth?code=...&user_id=...&email=...

Cursor:

cursor://manjuai.manju-ai/auth?code=...&user_id=...&email=...

The browser passes a short-lived code (not a JWT). The extension exchanges it for a Firebase id_token. If sign-in fails with certificate errors, sign out in the sidebar and try again.

Architecture

Extension (TypeScript + chat webview)
    │ HTTP 127.0.0.1:8770 (agent), :8772 (index)
    ▼
python/automate_agent (+ index_worker)
    │ HTTPS (default manjuai.tech/manju-api)
    ▼
Manju API — /manju-chat/local-agent/llm-step, auth, billing

Before each turn the extension checks that the configured API is reachable.

Settings

manju.backendUrl — API base (default https://manjuai.tech/manju-api)
manju.webAppUrl — sign-in site (default https://manjuai.tech)
manju.agentPort / manju.indexPort — local ports (8770 / 8772)
manju.serverLaunch — terminal (default) or extension
manju.serverTerminalFocus — focus Terminal when servers start
manju.pythonPath — Python executable (empty = python / python3 on PATH)
manju.defaultModel — signed-in default model id
manju.autoRetrieveContext — inject semantic search hits before each turn (default on)

Develop

cd manju-vscode
npm install
pip install -r python/requirements.txt
npm run check
pytest python/automate_agent/tests -q

Press F5 with the manju-vscode folder open (Extension Development Host).

Debug: set MANJU_LOG_SYSTEM_PROMPT=1 on the Python agent process to log the assembled system prompt.

License

MIT — see LICENSE.

ManjuAi