context-usage-mcp

A tiny Model Context Protocol server (TypeScript/Node, stdio) that exposes a single tool, get_context_usage, reporting the current session's raw token usage — so an agent can check after each task whether context has grown large and branch its behavior on an absolute threshold. Runs under both Claude Code and OpenAI Codex CLI, reading each host's own session transcript/rollout behind an auto-detected host-adapter layer.

What it does

Neither host hands an stdio MCP server the live token counts, so this server reads usage indirectly from the host's own on-disk session file, behind an auto-detected host-adapter layer. The output shape is identical across hosts.

Claude Code passes only CLAUDE_PROJECT_DIR (no session id, transcript path, or counts):

1. Derive the project folder from CLAUDE_PROJECT_DIR (non-alphanumerics → -).

2. Pick the most-recently-modified *.jsonl in ~/.claude/projects/<folder>/ as the current session (its transcript is being written while the tool is called).

3. Scan backward for the last assistant message carrying a usage block.

4. Sum input_tokens + cache_creation_input_tokens + cache_read_input_tokens → context_tokens.

The ~/.claude root can be relocated with CLAUDE_CONFIG_DIR.

OpenAI Codex CLI passes no project/session env var at all, but launches the server with its working directory set to the session's project dir. So the server uses process.cwd():

1. Resolve the Codex root (CODEX_HOME, else ~/.codex); rollouts live under <root>/sessions/YYYY/MM/DD/rollout-*.jsonl.

2. Scan recent day-folders (today + yesterday, widening to ~7 days), reading only each candidate's first session_meta line to match its cwd against process.cwd(); pick the freshest match.

3. Scan backward for the last token_count event and read its per-turn last_token_usage.

4. Map to the shared, disjoint breakdown. Codex's cached_input_tokens is a subset of input_tokens (verified in source), so input ← input − cached; then context_tokens = input + cache_read + cache_creation — no double-count.

Output is output_tokens + reasoning_output_tokens. The Codex model_context_window is ignored (raw counts only, strict parity with the Claude host).

Related MCP server: Claude Telemetry MCP

Install & run

Run the published package directly with npx — no clone or build required:

npx -y @beeltec/context-usage-mcp

This is the recommended way to wire it into a host (see the registration sections below). The -y flag skips the install prompt on first run.

Run from source (local dev) — for hacking on the server itself:

npm install
npm run build          # tsc → dist/, sets +x on dist/index.js
npm test               # parser unit tests

Then register the absolute path to dist/index.js (see the "from source" note in each registration section).

Host detection

The server auto-detects its host: CLAUDE_PROJECT_DIR set → Claude Code; else CODEX_HOME set or ~/.codex/sessions present → Codex; else it falls back to Claude. Set the environment variable CONTEXT_USAGE_HOST=claude|codex to force a host (the override wins over auto-detection).

Register with Claude Code (user scope)

claude mcp add --scope user context-usage -- npx -y @beeltec/context-usage-mcp

Verify with claude mcp list. Because MCP servers are loaded at session start, start a new Claude Code session before calling the tool.

From source: register the absolute path to this repo's built binary instead:

claude mcp add --scope user context-usage -- node /absolute/path/to/dist/index.js

Register with Codex CLI

Add an entry to ~/.codex/config.toml (root movable with CODEX_HOME):

[mcp_servers.context-usage]
command = "npx"
args = ["-y", "@beeltec/context-usage-mcp"]

Or use the CLI equivalent:

codex mcp add context-usage -- npx -y @beeltec/context-usage-mcp

Do not set a custom cwd for the server: Codex passes no project/session identifier, so discovery relies on the server inheriting the session's working directory via process.cwd(). Setting cwd breaks the cwd-match. MCP servers load at session start, so start a new Codex session before calling the tool.

From source: use command = "node" with args = ["/absolute/path/to/dist/index.js"] (or codex mcp add context-usage -- node /absolute/path/to/dist/index.js).

Distribution

Published to npm as @beeltec/context-usage-mcp and run via npx (see Install & run). Releases are published from CI with provenance on each GitHub Release; running from a local build is also supported for development.

The tool: get_context_usage

No arguments. Returns raw counts only — no percentage, no context-window detection. It may be unavailable early in a session before the first model response. Both a JSON text block and typed structuredContent (declared outputSchema) are returned with the same object. The shape is identical under Claude Code and Codex.

Available:

{
  "available": true,
  "context_tokens": 86026,
  "breakdown": {
    "input_tokens": 2,
    "cache_creation_input_tokens": 1001,
    "cache_read_input_tokens": 85023,
    "output_tokens": 225
  },
  "session_id": "…",
  "model": "claude-opus-4-8",
  "timestamp": "2026-07-20T14:17:50.751Z",
  "reason": null
}

Unavailable (never throws): numeric fields are null and reason explains why (no transcript/rollout found, empty file, or no usage-bearing message/token_count event yet).

Intended usage

The server returns raw numbers only; the agent decides what "too long" means. The pattern: call get_context_usage after each task and branch on an absolute token threshold defined in the agent's prompt/CLAUDE.md (e.g. wrap up or hand off when context_tokens exceeds a limit).

Known risks

• Internal file format. Both the Claude transcript JSONL and the Codex rollout JSONL are undocumented internal formats that may change between host versions; parsing them directly is an accepted fragility.

• Concurrent-session race. The freshest-file heuristic (freshest *.jsonl for Claude; freshest cwd-matching rollout-*.jsonl for Codex) is racy if multiple sessions run in the same project at once — it may pick another session's file.