Tokenmancer

See what every AI coding prompt really costs — a live AI-Credit meter, budgets, workspace analytics and a what-if simulator, inside VS Code.

Homepage: https://genesis-nexus.github.io/tokenmancer/

Every prompt you send fans out into a loop of model calls (plan → read → search → edit → verify). Tokenmancer reads Copilot's own agent debug log and lays each loop out as a priced, step-by-step timeline: which model ran each step, the context window growing call after call, and a cost bar splitting every bill into cache-read / cache-write / fresh-input / output — with output, a few hundred tokens, routinely the biggest slice.

Everything runs on your machine. Nothing is uploaded.

Setup

Tokenmancer reads a log Copilot only writes when you ask it to. Add this to your VS Code settings, then reload the window:

"github.copilot.chat.agentDebugLog.fileLogging.enabled": true

(An older github.copilot.chat.agentDebugLog.enabled setting is deprecated as of Copilot Chat 0.48 and does nothing on its own now — GitHub folded it into the setting above. If your Settings UI search hides it, edit settings.json directly, or run Preferences: Open User Settings (JSON) from the Command Palette.)

Send an Agent request, then open the Tokenmancer view in the activity bar. If the setting is off, the view says so and offers to jump you to it.

What you get

Live Meter — the current workspace's Copilot log, priced as it happens.

Budgets and alerts — set a monthly credit limit and get told at 50 / 80 / 100%. Spend is written to an append-only ledger, so month-to-date survives a restart; the gauge frames it against the calendar ("86% used · 14 days left in August · on pace to run out around the 20th"), because a seat's pool does not roll over. Alerts fire on a threshold crossing, never on a level, and are monotonic within a period — once 80% has fired, 50% cannot fire again — with a cooldown and an hourly cap behind that.

Status bar — ⚡ 42.0 cr · 1.4% at all times. Click it to open the meter.

Workspace Analytics — day-by-day trends, cache reuse, model-switching, context-pressure bands and plain-English recommendations across every session in a workspace.

Session Replay — browse past sessions on this machine by date and replay any of them.

What-if Simulator — exact o200k token counts, modelled across every rate row with cache and output levers.

Simple and Detailed

Every surface has a Simple / Detailed switch, and it starts on Simple.

Simple answers one question — what did that cost me, and why? Spend leads in dollars, each prompt gets a one-line explanation of where its money went, and step tables fold away behind a per-card control. A rotating Learn card introduces one idea at a time — why replies cost more than questions, why staying in one chat is cheaper, why the first message of a session is the expensive one — and each ends with a link into the exact detailed view that evidences it.

Detailed is the full instrument: per-step context bars, the four-way cost split, the every-number step table, and the complete analytics breakdown. Your choice is remembered and always beats the workspace default.

Commands

Settings

All under tokenmancer.* — pool size and default model, the monthly budget, alert cooldown and rate cap, privacy, and the default level of detail. Settings you leave alone fall through to ~/.tokenmancer/config.json, which a repo can also override with a committed .tokenmancer.json.

Privacy

Everything runs in the extension host; no data leaves your machine.

Your prompt text is shown — it is how you tell one loop from another, and it never leaves the extension host. Set tokenmancer.privacy.showPrompts to false before a screenshare and each loop falls back to an opaque ‹redacted a1b2c3› tag.

Shell commands and search strings are blanked by default (tokenmancer.privacy.showToolQueries), because a command line is where a token or connection string tends to get pasted.

Tokenmancer reports which files the agent read, never their contents, and absolute paths are withheld unless you ask for them.

Requires

VS Code 1.90+, and GitHub Copilot Chat with the agent debug log enabled (above).

MIT-licensed. Source and issues.