MCP Behavior 0.1.0

MCP Behavior 0.1.0

MCP Behavior 0.1.0 is the first public alpha. It runs strict YAML scenarios against an MCP server and checks returned values plus declared file, tree, SQLite, or custom-probe effects.

The release includes expectation tests, reviewed JSON baselines, and live candidate-versus-reference comparison. Stdio and Streamable HTTP targets use the official MCP Python SDK. Reports are available for terminals, JSON consumers, pull request comments, and JUnit readers.

The first-run demo requires no existing server or API key. It executes an intentional invoice regression: a response-only check passes, a declared file check catches the wrong total, and the corrected implementation passes. init starts with a file assertion; SQLite and response-only templates are also available. Human reports include escaped expected/actual previews, with complete JSON evidence retained.

The release also includes a reproduced historical CRLF regression in the official filesystem server. Published 2026.8.31 and 2.0.0 return identical successful tool content but write different line endings. The case credits Cliff Hall's upstream PR #5011, includes independent byte controls, and is exercised separately in CI. The 24-second demo replays actual stdout at a slower reading pace; its duration is not execution latency. A direct SDK test with the same byte assertion catches the regression too.

The clean-wheel gate executes ten example configurations, all starter commands, the invoice demo, and a semantic-digest repeatability check.

Install and try

Requires Python 3.11+ and uv:

uvx --from git+https://github.com/GWeale/mcp-behavior.git@v0.1.0 mcp-behavior demo

Alternatively, download the wheel below, verify it against SHA256SUMS, and install it with uv tool install ./mcp_behavior-0.1.0-py3-none-any.whl. This GitHub alpha is not a PyPI publication.

Compatibility

• Python 3.11 through 3.14 are tested on Linux, macOS, and Windows.
• Contract schema: 1
• Evidence schema: 1
• Baseline schema: 1
• MCP protocol client modes: auto, 2026-07-28 modern, and SDK legacy

Security boundary

Servers and configured commands run as trusted code. The release bounds output and process lifetime, contains relative observer paths, opens SQLite read-only, rejects credential-bearing target URLs, avoids HTTP redirects, and sanitizes request and response evidence before persistence. These controls do not isolate hostile code.

Artifacts

The release contains a wheel, source archive, SHA256SUMS for those two packages, the demo MP4, original asciinema cast, captured stdout, sanitized case receipt, and a Markdown state report. The workflow builds packages from the tagged commit and runs the complete release gate before publishing this GitHub prerelease.

Local candidate validation: 84 tests passed with 87.05% coverage; the complete release gate and clean-wheel smoke passed. The product PR's 15 CI checks passed, including Python 3.11–3.14 on Linux, macOS, and Windows, packaging, and the official-server case.