Context Control

by T&N Control

A VSCode / Cursor / Windsurf extension that monitors AI coding context usage of your local AI assistant conversations (Claude Code, Codex, Cline) and warns you before the context window fills up — then generates a portable handoff markdown you can paste into another AI session.

Everything runs 100% locally. No conversation ever leaves your machine.

Future provider/editor work—including Gemini, Cursor-native conversations, Antigravity, and other AI coding tools—is tracked in the product roadmap.

What it does

• Reads conversation history from local storage (no API keys, no network).
• Tracks token usage against the model's real context window.
• Shows an explicit live context reading (CC Ctx: 45%) and, when reported, a separate provider-quota reading (CC Quota: LIMIT · Ctx: 76.1%).
• Warns at a configurable threshold, alerts when critical.
• Predicts how many messages remain until critical, and estimates cost.
• A Cockpit side panel with live gauges: context usage and your provider quota (Codex 5-hour & weekly limits, with a live reset countdown) — no login or API key, read straight from local files.
• A Dashboard charting every session across tools: ranked context meters with the warning/critical thresholds drawn on the track, provider-quota meters with a live reset countdown, and estimated cost by project — with the full sortable table underneath.
• Generates a structured handoff (.md) with the initial/latest request, model-switch history, decisions, TODOs, referenced files, and a full-fidelity timeline of every meaningful user/assistant turn—without trimmed gaps.
• Keeps Generate Handoff scoped to the active project. In a multi-root window, it asks which project to use when the active editor does not identify one.

Token counting

• Claude Code exposes real token counts in each message's usage field, and Codex reports its real model_context_window — Context Control uses those directly (most accurate).
• For sources without usage data, it falls back to a characters-per-token estimate. (Up to 0.2.2 this used tiktoken, which was itself only an approximation for Claude models while accounting for ~11MB of the download.)

Session files are read incrementally — each refresh reads only the bytes appended since the last one, so tracking a 50MB conversation costs about two milliseconds per turn rather than a full re-parse.

Install

VS Code — open the Extensions panel (Ctrl+Shift+X), search “T&N Context Control”, and click Install. Or run:

code --install-extension tn-control.tn-context-control

Windsurf / Cursor — install the packaged file: Extensions panel → … → Install from VSIX… → pick tn-context-control-<version>.vsix. (An Open VSX listing for in-app search is planned.)

Getting started

1. Install the extension and reload the window.
2. Open a folder where you use an AI coding assistant (Claude Code / Codex / Cline).
3. A status bar item appears bottom-right, e.g. CC Ctx: 45% 116k/258k — Ctx is your current context-window usage, not your account quota. Hover it for tokens/message, an estimate of how many messages remain until critical, and cost.
4. It turns yellow at the warning threshold and red when critical.
5. Run a command (Ctrl+Shift+P → type “Context Control”):
   • Open Dashboard — every session across tools, sorted by usage.
   • Generate Handoff — writes a portable summary .md to .ai-memory/ so you can continue the work in a fresh AI session before you run out of context.

Develop from source

npm install
npm run bundle   # production bundle (or: npm run compile)
npm test         # run the unit tests

Press F5 to launch an Extension Development Host.

Commands

Settings

Supported adapters

Known limitations

• Where a provider does not report real usage, token counts are a characters-per-token estimate.
• The Cline adapter is unverified — field names/paths may differ per version.
• The handoff generator is rule-based (regex/keyword), not LLM-powered, so its summaries are best-effort.
• Context window is inferred per model (Opus/Sonnet 4.5+ → 1M, Haiku/older → 200k) unless the provider reports it (Codex). If usage exceeds the inferred window, the limit is escalated so usage never reports a nonsensical >100%.
• Cost is estimated from published Claude prices (Opus $5/$25, Sonnet $3/$15, Haiku $1/$5 per 1M; cache read ×0.1, write ×1.25–2.0). GPT/other providers are not priced.
• Live session-quota percentages are available from Codex. Claude Code does not expose the percentage leading up to its limit, but its local transcript records a rejection when the 5-hour or weekly limit is actually hit; Context Control shows that as Quota: LIMIT and clears it after reset or a successful response. Context and quota are always labelled separately.

Reporting a problem

Run Context Control: Show Logs and include the output — it records adapter failures, scan timings, and watcher errors. Bugs can be filed through GitHub Issues or the extension's Marketplace Q&A tab.

Privacy

Context Control reads local files only and never transmits conversation content. docs/research/findings.md and the .ai-memory/ output folder are gitignored.

License

Source-available proprietary — © 2026 T&N Control. The code may be viewed and forked through GitHub, and official releases may be used as described in LICENSE; modification, redistribution, and commercial exploitation require prior written permission. This is not an open-source license.