Skip to content
| Marketplace
Sign in
Visual Studio Code>Machine Learning>OneFirewall AINew to Visual Studio Code? Get it now.
OneFirewall AI

OneFirewall AI

OneFirewall

|
15 installs
| (1) | Free
OneFirewall AI coding agent for VS Code (powered by the onefirewall-ai CLI)
Installation
Launch VS Code Quick Open (Ctrl+P), paste the following command, and press enter.
Copied to clipboard
More Info

OneFirewall AI for VS Code

A coding agent in your editor, powered by the onefirewall-ai CLI. The extension is a thin client: it starts onefirewall-ai --stdio in the background and renders the conversation, so the agent behaves exactly like it does in your terminal (same tools, sessions, permission rules, MCP servers and project settings).

Requirements

  • VS Code 1.95 or newer

  • Python 3.10+ with the CLI, version 0.3.0 or newer (0.2.x still works, without agents):

    pip install -U onefirewall-ai
    

The extension finds the CLI automatically (onefirewall-ai on your PATH, then py -3 -m onefirewall_ai on Windows, python -m onefirewall_ai, python3 -m onefirewall_ai). If it lives in a virtual environment, set onefirewall.cliPath, for example C:\venv\Scripts\python.exe -m onefirewall_ai.

Features

  • Chat with an agent that works in your workspace: it reads and searches code, edits files and runs commands.
  • Approvals you control: every edit or command can be reviewed first. Edits show a colored diff, with Open diff for VS Code's native diff view. Choose Allow, Always allow or Deny, and optionally tell the agent what to do instead.
  • Permission modes (pick from the mode button in the input, or press Shift+Tab to cycle): Ask before changes, Edit automatically (edits without asking, asks before commands), Auto (runs edits and safe commands without asking and pauses only for risky actions such as git push, recursive deletes, publishing, cloud changes or sudo, telling you why), Plan (research first, then approve a plan) and Bypass permissions.
  • @ file mentions: type @ to attach workspace files to your message.
  • Editor awareness: the file you have open appears as a chip in the input (prompt.ts, or prompt.ts:12-20 with a selection) and is sent with your message, together with the selected text and your open tabs. The agent reads the file only when it is relevant. Click × on the chip to leave it out of the next message; it comes back when you switch files. The agent also sees VS Code's diagnostics (errors and warnings) after it edits a file.
  • A focused conversation view: each new question moves to the top while the answer fills in below it. Tool calls, sub-agents, reasoning and approvals are grouped in one collapsible activity row ("Worked for 1m 12s · 8 steps"), with an animated OneFirewall mark while the agent works, and only the final answer is shown in full.
  • Always know what it's doing: while the agent works, the last line of the conversation shows the animated OneFirewall mark and one word for what it's doing ("Editing main.py…", "Running tests…", "Cooking up a plan…"). Hover it for elapsed time and steps. Scroll away and a small pill shows it and jumps back; VS Code's status bar shows it too.
  • Keep typing while it works: messages are queued as gray bubbles and the agent reads them at its next step. ↑ edits queued messages, × drops one, Ctrl+Enter sends right away, and Esc stops the current work and sends what you queued. (Needs CLI 0.2.1+.)
  • Model picker in the input footer: click the model name (or type /model) to switch workspace or model.
  • Agents working in parallel (see below), with a quiet counter in the input footer and an Agents panel.
  • Slash commands, like the terminal CLI: type / for /new, /history, /model, /mode, /plan, /compact, /context, /cost, /todos, /undo, /init, /agents, /mcp, /terminal, /login, /logout and /help.
  • MCP servers: see each server's status and tools, reload them, add a server (local command, HTTP or SSE), edit the user or project config, and trust servers defined by the project (/mcp, or Manage MCP Servers).
  • History with generated chat titles, search and delete. Sessions are shared with the CLI; Continue Chat in Terminal (/terminal) opens the current conversation in the terminal.
  • Task list above the input, and Undo last turn to restore the files the agent changed in its last turn.
  • Commands: Explain / Fix Selected Code (editor context menu), Start Agent Task, New Chat, Stop, Resume Session, Switch Workspace / Model, Set Permission Mode, Undo Last Turn, Show Agents, Manage MCP Servers, Add MCP Server, Continue Chat in Terminal, Show Logs, Login / Logout.

Agents

The agent can hand parts of a task to agents that work in parallel and in the background: built-in explore (read-only search), plan (read-only, returns a plan) and general, plus your own agents defined as Markdown files in .onefirewall/agents/ (project) or ~/.onefirewall-ai/agents/ (user). Agents never get more permissions than the chat's mode. Needs CLI 0.3.0 or newer.

  • In the conversation each agent is one row in the activity list: its description, a small badge with its name and, on the right, what it is doing right now ("Reading parts.tsx…", "Searching for “session”…") or, once done, "12 steps · 1m 4s". Expand the row to see its latest steps. Rows keep updating after the main agent moves on.
  • Approvals an agent asks for show which agent is asking, at the top of the card.
  • Results: the main agent tells you which agents are working and finishes its reply; under it, "● 2 agents are working in the background · Stop all". You can keep chatting. When an agent finishes, its report comes back to the main agent as a new reply headed by a small chip ("“Security audit” finished"), where it tells you what was found; after the last one it gives the complete answer. If the main agent is still working, it gets the report before its next step ("Report from “…”" in the activity list).
  • Stop (Esc) stops the main agent only; agents keep running and still report back.
  • Footer counter: "● 3 agents" next to the model button while agents run (the dot turns amber when one needs your approval; at narrow widths it shows "● 3"). Click it to open the Agents panel: the main agent and every agent of this chat, with time, steps and tokens. Click an agent for its brief, report (with Copy), steps, Stop, Run in background and Open transcript (a read-only Markdown tab). Also via /agents, the ⋯ menu or OneFirewall: Show Agents.
  • @ mentions: type @ to pick an agent by name (@explore) as well as files. VS Code's status bar shows "OneFirewall AI · 3 agents working" while agents run.

Sign in

Run OneFirewall: Login (or click Sign in in the chat). Sign in with your OneFirewall account in the browser, paste an API key, or connect any OpenAI-compatible provider (OpenAI, Groq, Ollama, LM Studio, Azure OpenAI…). The API key is kept in VS Code's secret storage and handed to the CLI via environment variables. It is never written to disk by the extension.

Settings

Setting Default Description
onefirewall.cliPath "" Command used to start the CLI (auto-detected when empty). May include arguments.
onefirewall.permissionMode default Mode the agent starts in: default, accept-edits, auto, plan or bypass. Picking Ask, Edit automatically or Auto in the chat updates it.
onefirewall.autoOpenDiff false Open VS Code's diff editor automatically when the agent proposes an edit.
onefirewall.spinnerVerbs {"mode": "append", "verbs": []} Words shown while the agent thinks; append adds to the built-in list, replace uses only yours.
onefirewall.spinnerTips false Show a one-line tip under the working indicator.
onefirewall.includeEditorContext true Send the active file, selection and open tabs with each message.

Troubleshooting

Run OneFirewall: Show Logs to see the CLI's output. If the agent process crashes, the extension restarts it (up to three quick restarts) and resumes the current session.

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