FutureX for VS Code

The extension is published on the Visual Studio Marketplace as futureim.futurex. Search FutureX in the Extensions view, or run either of these:

code --install-extension futureim.futurex

ext install futureim.futurex

It requires Visual Studio Code 1.90 or newer on Windows, macOS, or Linux. There is nothing else to install — no API key to paste, no runtime to download.

The extension declares no activation events, so it stays off your startup path entirely: it loads the first time you open the panel or run one of its commands.

Click the FutureX icon in the activity bar, or press Ctrl+Alt+X (Cmd+Alt+X on macOS). Type a question about the file you have open and press Enter.

You do not need to paste any code. The panel attaches the file path and language, the code itself — the lines you selected, or the head of the file when you have selected nothing — and any errors or warnings the language server is reporting on them. That context block is shown in the transcript, collapsed, so you can read exactly what was sent.

The first 5 prompts run without an account. The counter lives on the server, and the panel footer shows what is left.

Run FutureX: Sign In. Your browser opens the FIM portal with a short code already on your clipboard; approve it there and the extension is signed in. There is no API key to paste and none to leak.

Credentials are stored in your operating system's keychain through VS Code's secret storage — Credential Manager on Windows, Keychain on macOS, libsecret on Linux. They are never written to a settings file and never synced.

Signing in unlocks the full agent (reading the project, editing files, running commands) and the Pro model. A FIM account is free to create.

Every command is in the command palette under FutureX. The code commands are also on the editor's right-click menu, under a FutureX submenu.

Two dropdowns sit at the top of the panel: which model answers, and how long it thinks first.

• Pro — the hardest problems, deepest reasoning, largest answers. Requires a signed-in account.
• Fast — the everyday workhorse, and the default.
• Light — light reasoning, quicker turnaround.
• Eco — cost discipline: short, direct answers.

Each model allows a different set of thinking levels, and the picker only offers what the selected model supports. The gateway enforces the same rule server-side, so a model can never run at an effort it does not allow. Every model bills at the same flat rate of $1 per 1M tokens, input and output alike.

Reads are never gated — they are how the agent answers anything. Everything that changes something stops and asks, with the change itself in front of you: the unified diff for an edit, the literal command line for a terminal run.

Consent is per kind, not global: agreeing that the agent may edit files says nothing about whether it may run commands. The four kinds are edit, create, delete and run.

How long a grant lasts is set by futurex.approvals:

• ask — every time. The default, and a grant covers one call.
• session — until the window closes.
• workspace — remembered for this workspace.

FutureX: Reset Permission Approvals forgets all of them. In a workspace you have not trusted, the agent drops to read-only tools: no edits and no commands, regardless of these settings.

Nothing is uploaded wholesale. There is no indexing pass, no background scan, and no copy of your repository on a server.

• With each prompt: the open file's path and language; the code itself — your selection (up to ~6,000 characters), or the opening ~4,000 characters of the file when nothing is selected; the diagnostics on it; and a git summary (branch, upstream, which files are staged, unstaged or untracked, and your origin remote URL). This block is printed in the transcript, collapsed, so you can read all of it.
• Once per session: a short project card (name, language, framework, package manager, test runner — read from your manifests) and a machine summary (OS, shell, which tools are on your PATH). No file contents. Because it is sent once rather than per prompt, it cannot live in the transcript; it is written verbatim to the output channel instead (FutureX: Show Logs).
• On demand: once signed in, the agent reads specific files when it needs them. Those reads happen inside your editor, confined to your workspace folder; only the result goes back.

Each of the three is switchable: futurex.context.activeFile, futurex.context.git and futurex.context.project.

All settings live under the futurex. prefix. Open them with Preferences: Open Settings and search “futurex”, or edit settings.json directly.

Both futurex.model and futurex.effort are window-scoped, so a workspace settings.json can override either for a single project.

The panel is blank after an update

Reload the window (Developer: Reload Window). Installing a new version over a running one tears down the old extension host, and the panel's content lives there — so the view frame remains while its contents do not. This affects every VS Code extension, not just this one; closing and reopening the panel does not help, because the provider lives in the extension host.

A command reports “command not found”

That means the extension failed to activate. Open FutureX: Show Logs and the Developer: Show Logs… → Extension Host output, and send us what they say. Reloading the window fixes the common case (an update applied while the old version was running).

The free prompts are used up sooner than expected

The allowance is counted per install and, more loosely, per network. On a shared office connection the network bucket can be reached by somebody else first — the message will say so. Signing in with a free FIM account clears it either way.

It cannot reach FutureX

The extension talks to https://api.futureim.org. Behind a proxy or on a self-hosted FIM deployment, point futurex.apiUrl at the right origin — the value is the bare origin, with no /api/v1 suffix.