Skip to main content

Context Audit CLI

Feature row 58 — Sprint CA-OSS (Partial)

Phase 1 ships today: the CLI, JSON output, and a handful of common MCP client configs. Phase 2 (README polish, CI), Phase 3 (PyPI), and Phase 4 (PDF output, web UI) are in progress.

What this is​

behavry-ca (Context Audit) is a standalone CLI that answers a question every agent user eventually asks: how many tokens are my MCP integrations costing me before I've even typed a prompt?

The tool parses the config files of the common MCP clients — Claude Code, Claude Desktop, Cursor, VSCode MCP extensions — identifies every registered server, calls each server's tools/list, counts the tokens in the raw response, and prints a breakdown by server and by tool.

You don't need Behavry's backend to use it. You don't need a tenant. It's a single binary (or pip install behavry-ca) and a few seconds of runtime.

Why it matters​

Typical findings after running behavry-ca on a real developer setup:

  • A single MCP server registering dozens of rarely-used tools, eating 15k+ tokens of context before any work starts
  • Descriptions carrying multi-paragraph usage examples that dominate the token count
  • Multiple servers duplicating similar tools (two different "filesystem" servers both exposing read_file)

The output is the same data Context Window Governance uses in the commercial product — only without the enforcement / policy layer.

Usage​

behavry-ca scan

By default, scan auto-detects configs in:

  • ~/.claude.json
  • ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
  • ~/.cursor/mcp.json
  • ~/.vscode/mcp.json

Point it at a specific file with --config:

behavry-ca scan --config /path/to/claude_desktop_config.json

Output:

Server                 Tools   Tokens  Per-tool avg
filesystem             12      3,412    284
github                 48      14,882   310
slack                  21      8,199    390
brave-search           3       612      204
---------------------------------------------------
TOTAL                  84      27,105

Add --json for structured output, --format table|csv|md for alternatives.

Recommendations​

behavry-ca recommend runs a scan, then applies a set of rules of thumb and suggests:

  • Candidates for removal — servers with no tools that have been called in the last N days (requires audit log access)
  • Candidates for trimming — tools whose descriptions are the top 10% by token count
  • Duplicates — tools with similar names and signatures across different servers

Rules live in recommendations.yaml and are customizable.

JSON output​

The JSON shape matches the commercial product's Context Gate metrics, so teams that start with the OSS CLI can hand the output directly to a Behavry Professional / Enterprise trial for ongoing enforcement.

Source and downloads​

  • Repo: github.com/behavry/behavry-ca
  • Install:
    • pip install behavry-ca (when Phase 3 ships)
    • brew install behavry/tap/behavry-ca (planned)
    • Signed binary releases: macOS arm64 / x86_64, Linux arm64 / x86_64
  • License: Apache 2.0