AI subscription management and usage for VS Code

The extension also bundles an experimental CLI BYOK bridge that offers your existing Codex/Claude subscription logins as models in the Copilot model picker. The bridge is still in development and not fully working yet. Its settings are tagged Experimental in the Settings editor, and it can also run standalone as a local model endpoint.

Run several AI subscriptions without losing track of them. AI Usage switches Claude Code and Codex between saved accounts without signing in again, and shows live rate-limit and quota usage for Claude Code, Codex and GitHub Copilot in the status bar and, in the Agents window, behind a chip beneath the chat input.

• Status bar: 17% (3h) 25% (3d) for Claude, where parentheses show the time until reset—not the fixed quota-window length. The item turns yellow at 80% and red at 95%.
• Chat chips: compact percentages such as Claude 17% 25% beneath the chat input for the agent the current chat is using; click for the named windows and reset countdowns.
• Per-chat tokens: 400k tokens for the newest Claude or Codex chat in the current workspace; click for the exact input, output and cached-input breakdown.
• Details panel: click a chat chip or the Copilot status bar item, or choose Back in an Accounts menu, for plan, account or organization, reset countdowns and source, with Refresh, Open log and Settings actions.
• Organization aware: Copilot follows the GitHub account whose Copilot organization owns the workspace repository, so org-billed seats show org data.
• Gentle on the services: one shared cache for all open windows, per-service check intervals and Retry-After aware backoff. Codex's sessionLog source reads usage offline; earned-reset availability requires an app-server check; auto does not add one after a successful cheaper reading.
• Honest when things fail: a failed refresh keeps the previous reading and greys it out after 15 minutes.

Nothing is shown for a tool that is not installed or signed in. Usage checks only read the tools' own login files. Automatic profile rotation is on by default and acts once at least two profiles are saved. Earned Codex rate-limit reset redemption is enabled by default for a limited saved account when Codex reports available credits. Enable aiUsage.codex.autoReset.confirmationRequired to require editor approval before spending a credit; without an approving editor, confirmation mode leaves it untouched. Keep-alives and rotation can make background model calls and refresh tokens; rotation also changes the native login.

With the default icon-only labels, the status bar stays compact while showing each live reset countdown:

Labels are configurable; iconAndName adds provider names without changing the usage figures:

With several saved accounts, the active profile's number in the Accounts list precedes the figures (aiUsage.<claude|codex>.statusBar.accountNumber); here Claude runs on profile 3 and Codex on profile 4:

High usage is deliberately visible: an item turns yellow at 80% and red at 95%. Here Claude's 92% window resets in 3 hours while Codex remains neutral at 77%. Countdown labels always use one largest unit (42m, 3h, or 4d):

Hover a status item for its named limit windows, plan and reset times, the source of the reading and the active authentication profile. Copilot also identifies the account and, for organization-billed seats, the organization and its premium-request usage:

Claude	Codex	Copilot

Click a chat chip or the Copilot status bar item, or choose Back in an Accounts menu, to open the AI Usage menu of all services: each provider's windows with their reset countdowns, the source of every reading, and the refresh, log and settings actions in one picker. A service's Accounts row opens its Accounts menu:

Authentication profiles

Run AI Usage: Manage Claude/Codex Authentication Profiles from the Command Palette, click the Claude or Codex status bar item, or choose a service's Accounts row in the AI Usage menu, to open that service's Accounts menu. You can save and name the current login, import a credential JSON file, export every saved profile to a file and import it on another computer, and switch among up to 20 profiles per service.

Codex: experimental. Switching Codex accounts (by hand or by automatic rotation) and the Codex account proxy are experimental: open Codex chats and background processes can keep the previous login or fail a turn until they are restarted, and the behaviour can change between releases. Reading Codex usage works normally, but keep aiUsage.codex.checkIntervalMinutes and aiUsage.codex.api.minAccountIntervalMinutes long enough that the checks do not run into the provider's rate limits. Claude account switching is not affected. A profile is private, kept by the account service on this host, or a project profile, kept with its login in the workspace folder's .ai-usage.profiles.json and listed whenever that folder is open; see Private and project profiles. Every saved profile is listed with its login email, its last usage reading and check time, and the active one is marked:

Saved copies live with the account service in ~/.ai-usage/profiles.json (mode 0600), or in a project's profile file, never in workspace settings or the extension's ordinary storage. Activating a profile atomically updates the native file already shared by the CLI and vendor extension (~/.claude/.credentials.json / $CLAUDE_CONFIG_DIR, or ~/.codex/auth.json / $CODEX_HOME). The file is forced to mode 0600 on Linux/macOS and inherits the user-profile ACL on Windows. Claude MCP credentials in the same file are preserved. This native active copy remains plaintext because Claude Code and Codex require that format; only one selected profile is exposed there at a time. For Claude, activation also updates the token's account UUID/email in ~/.claude.json and clears its account-bound caches, so /status and /usage do not retain the previous Team member's identity.

A convenient setup is: sign in with the CLI, save the current login as a profile, sign in with the next account, and save again. An imported profile is saved but not activated until you choose it. Claude reads its credential file per turn, so open Claude chats and CLI sessions use a switched login from their next turn, with no restart. Codex keeps its login in memory, so open Codex chats follow a switch only through the optional Codex account proxy (aiUsage.codex.proxy.enabled), which routes their requests through AI Usage and attaches the active login per request; without it, AI Usage offers an extension-host restart. The saved profiles and their logins are kept by the account service on the host the extension runs on (the remote in a Remote-SSH, WSL or container window), so every window connected to that host and the ai-usage command there share them.

Save current login… offers to create a new profile or to replace one that is already saved, which is also how a profile whose token has expired is repaired after signing in with that account again:

Manage saved profiles… contains Import credential JSON…, Export or import saved profiles…, sign-in, rename, move, disable/enable and delete. A disabled profile is left out of automatic rotation, keep-alives and background usage reads; it stays saved with a Disabled label, every account row has a button to disable or enable it, and it can still be chosen by hand. The transfer item leads to both actions. Export saved profiles… writes the saved Claude and Codex profiles, logins included, to a JSON file and opens it in the editor, and Import saved profiles… reads it on another computer: profiles not saved there are added, and a saved profile whose login is missing gets it back. Nothing is activated. The file holds live login tokens in plain text, so delete it once imported, and mind that a copied login is the same session on both computers; see Moving profiles to another computer.

The account service and the ai-usage command

The account service owns usage collection, saved credentials, account rotation, earned resets, the Codex proxy and the CLI bridge. VS Code displays its readings and forwards settings and user actions. Installed in the background, the service keeps working when VS Code closes. Otherwise one editor window hosts the same service and other windows share it; both modes keep profiles in ~/.ai-usage/profiles.json. Existing nonconflicting VS Code profiles are migrated there on connection.

Engine settings are available through ai-usage config without the aiUsage. prefix. Persisted engine settings are authoritative when an editor connects; explicit editor and CLI edits update the shared configuration. Presentation settings stay local to each editor. Enable aiUsage.claude.advanced.rotationDiagnostics or aiUsage.codex.advanced.rotationDiagnostics to see every saved account's rotation score and the reason the active account is kept in its status tooltip.

See service ownership and deployment modes, configuration, and the optional Docker deployment proposal. AI Usage: Account Service… installs, upgrades, starts, stops or removes the background service. It requires Node.js 20 or newer; the bundled CLI bridge requires Node.js 22 or newer. VS Code's runtime can be used when available. AI_USAGE_NODE selects the runtime explicitly.

The service comes with the ai-usage command, on the PATH of every VS Code terminal (add ~/.ai-usage/bin to your own PATH for other terminals). ai-usage alone opens a live view of both services with keys to switch, send keep-alives, run a rotation sweep and toggle keep-alive or rotation; ai-usage status, list, use, save, login, keepalive, rotate, export, import-profiles, config, service and log do the same from arguments, with --json where it helps, and --project names a project folder whose profiles to list or save into. Engine settings changed with ai-usage config are reflected in connected VS Code windows. Details: Account service.

AI agents: the MCP server (experimental)

An AI agent can do the same through the Model Context Protocol: ai-usage mcp is a stdio MCP server whose tools list saved profiles and native-login usage, including unsaved accounts, with freshness and model-window limits, request a reading for a saved or native account and, when you allow it, switch the active account or run a rotation sweep. It is off by default (aiUsage.mcp.enabled); while it is on, agents in the VS Code window see it as AI Usage accounts without any configuration. Choose Set up MCP server… in the top-level AI Usage menu to enable the bundled server, then register it with the Claude or Codex CLI for terminal sessions. The CLI registration also appears in that service's Accounts menu. aiUsage.mcp.switching decides whether agents may switch at all. Details: MCP server for AI agents.

Account keep-alives and automatic rotation

Click the Claude or Codex status bar item to open that service's Accounts menu (or choose Accounts in the details panel); Back leads to the AI Usage menu of all services. This menu lets you switch accounts manually and save the current login to a new or existing profile. The ▷ button on a saved account's row sends it a keep-alive right away, and Manage saved profiles… → Send keep-alive now… sends one to a chosen account or to all of them. Account keep-alive and usage collection (aiUsage.<service>.keepAlive.enabled) and automatic account rotation (aiUsage.<service>.autoRotate.enabled) are set in Settings, independently per service (both are on by default, keep-alive with the smallest model); the menu shows what each one is set to. The active login is never kept alive or checked from its profile home; it is read through the native login it is in use with. Rotation chooses the next account from stored readings and never sends a keep-alive; aiUsage.<service>.autoRotate.updateUsageBeforeRotation (off by default) reads the target's usage right before the switch.

• Claude sends what is date today using haiku every 2 hours per saved account, in ~/.claude-tmp.
• Codex sends the same small request every 6 hours per saved account, using gpt-6-luna in ~/.codex-tmp, then reads account limits through its CLI. Its model is configurable; an empty model setting uses the CLI default.
• Both collect usage for inactive accounts. The authentication profile list shows each account's last usage, check time and errors. Model calls consume subscription usage and run while the shared runtime is running.
• Automatic rotation switches as soon as the active account's usage reaches a threshold (usage ≥ threshold), or would at its measured usage rate before the account can be read again; the active account is read the more often the faster it approaches a threshold. Claude has two: 95% for the 5-hour window and 99.5% for the weekly windows (all models, and 7d Fable when it counts). Codex's weekly threshold is 99%; a reported Codex 5-hour window rotates only once used up. The account switched to must be below the thresholds in every window. Both services offer soonest reset, even pace, least waste and saved order (autoRotate.strategy); Claude defaults to soonest reset and Codex to saved order. Both can also switch proactively to a clearly better account (autoRotate.trigger). Codex waits when its active quota naturally resets within five minutes and skips a candidate about to reset. When a Codex account is limited, aiUsage.codex.autoReset.enabled can instead redeem one of its provider-reported earned reset credits, preferring rotation unless no account qualifies or a credit would expire first. The Codex tooltip and profile rows show the remaining credits and expiry when known. Rotation and reset redemption work without keep-alives. Details: Rotation strategies and thresholds.

Each profile row carries the outcome of its last check: usage per window with the reset countdown, the check time, and a warning when the keep-alive or the usage check failed. An account whose login failed (expired or revoked) is marked Login problem; choosing it opens a small menu instead of activating it: Renew the login… runs the CLI login in a terminal with a separate home inside the keep-alive home and stores the new login in the profile, Try a keep-alive sends a keep-alive, which refreshes an expired token and, when the login cannot be refreshed, offers Sign in again in its failure notification, Select anyway activates the login as it is, and Back returns to the accounts. The same sign-in is available for any profile through the sign-in button on its row and as Sign in again… under Manage saved profiles…; either way the active login is replaced only when that profile is the active one.

Usage history. Every reading of every saved account, every switch (by hand, or by rotation with the accounts it considered and why it did or did not pick them), the sweeps that switched nothing and the stretches with every account at its limit are kept by the service for a year in one file per month, so you can see how much of each account you use and how well rotation works. AI Usage: Show Usage History… (also Usage history in the details panel) and ai-usage history summarize a period, with an estimate of how many accounts your weekly use needs, and export CSV or JSON Lines. Settings: aiUsage.history.*; file format and figures in Usage history.

Models, keep-alive periods, rotation thresholds, CLI paths, dedicated homes and usage sources are configurable under aiUsage.claude.* and aiUsage.codex.*; see account automation settings. Background checks swap credentials only inside the dedicated home, save refreshed tokens back to the profile, and remove the staged credential file afterward. For a checked active account, refreshed tokens are also written back to its native login when that login has not changed during the check. The account service runs the schedule, whether VS Code is open or not, and persists it across restarts.

Authentication profile examples

The Codex Accounts menu marks the current account as Active and offers Save current login…, Manage saved profiles…, account features and settings. Profile order can be changed from the management submenu or with ai-usage move <service> <profile> up|down.

After activation, the AI Usage menu names the selected profile beside Accounts, shows its detected plan on the right, and follows with its usage and the source of the reading:

Back at the end of an Accounts menu returns to the AI Usage menu of all services:

Installation

1. In VS Code open the Extensions view (Ctrl+Shift+X / Cmd+Shift+X), search for AI Usage, click Install. Or install from the command line:

   code --install-extension p0l0us.ai-usage-vscode-plugin
   

2. Sign in to the tools you use, if you have not already: run claude once, run codex login once, and sign in to GitHub in VS Code for Copilot.

3. The usage items appear in the status bar within a few seconds. For Copilot, VS Code shows one dialog asking to let AI Usage use your GitHub account; click Allow.

Requires VS Code 1.137 or newer. Works in Cursor and VSCodium from Open VSX.

Remote development (SSH, WSL, containers, tunnels)

The extension runs where your AI tools are signed in. In a Remote-SSH, WSL, dev container or tunnel window it is installed and runs on the remote machine, so:

• It reads the Claude and Codex logins on the remote machine. If you use those CLIs locally instead, they will not appear; install the extension locally too, or sign in on the remote side.
• Copilot uses the GitHub account VS Code is signed in with; that works in remote windows as usual.
• Settings for sources and intervals belong in Remote Settings, not the local user settings.
• Authentication profiles belong to that extension host's account service too. A Remote-SSH window manages the remote machine's profiles; a normal Windows/macOS/Linux window manages the local machine's profiles.
• The Agents window is the exception: it runs only local extensions, and blocks extensions with code until you allow them. The chip there needs a local install plus a one-time allow step; run AI Usage: Agents Window Setup Guide from the Command Palette or see docs/AGENTS_WINDOW.md.

Settings you are most likely to touch

Existing saved source modes and check intervals remain unchanged; new Claude and Codex configurations use auto and 30-minute checks.

The full list is in docs/CONFIGURATION.md. Where the numbers come from, how caching and backoff work and how the chat chip is built: docs/INTERNALS.md.

Troubleshooting

• A service is missing: it is not signed in on the machine the extension runs on (see the remote note above). Output → AI Usage shows what was found; Open log in the AI Usage menu opens it:

  

• A chat chip shows n/a: the chat's agent is not signed in where the extension runs. In the Agents window that is your local computer even for remote sessions; sign in there with the same account (claude once, or codex login) and the figures appear. Do not use aiUsage.chatChips.debug for this: it shows every service's chip, including Copilot's in a Claude chat.

• Copilot shows “connect”: click it and allow access to your GitHub account. The AI Usage menu offers the same action:

  

• Numbers are grey: the last refresh failed; the tooltip says why. Rate limits clear on their own.

Contributing

Bug reports, ideas and pull requests from any developer are welcome. Open an issue at https://github.com/p0l0us/ai-usage-vscode-plugin/issues (include the VS Code version and the relevant lines from Output → AI Usage; tokens are never logged) or send a PR. See CONTRIBUTING.md for the workflow and docs/PUBLISHING.md for building, dev-installing and releasing.

Most wanted: additional providers (Gemini CLI, Cursor, …), a Claude source that does not depend on the undocumented OAuth endpoint, tests.

License

GPL-3.0-only. See LICENSE.