SpanHound 0.6.1 — package documentation and first-run feedback

This documentation and packaging update makes the PyPI project page easier to
use: demo images, the video, and documentation links now resolve from package
pages as well as GitHub. It adds a first-run feedback form and links to the
Agent Observatory terminal/browser dashboard.

Try the recorded regression

uvx spanhound demo

The bundled scripted trace shows a blocked shell operation and its recorded
diagnostic and approval evidence. No API key or MCP client is needed for the demo.

For a supported Python agent, install in the agent's environment:

python -m pip install 'spanhound[auto]'
spanhound run --trace-dir ./traces -- python agent.py

The core capture and query behavior is the same as 0.6.0. The PyPI package,
MCP Registry metadata, and desktop bundle are versioned together at 0.6.1.

48-second walkthrough · User guide · First-run feedback

SpanHound v0.6.0 — one-command Python capture

SpanHound can now capture supported Python agents in one command, without manually setting OTLP exporter variables or adding tracing code to the agent.

In the agent's Python environment:

python -m pip install 'spanhound[auto]'
spanhound run --trace-dir ./traces -- python agent.py

The runner starts a loopback receiver, configures OpenTelemetry export for the child process, and stops the receiver when the program exits. The optional package includes instrumentation for OpenAI, Anthropic, OpenAI Agents, HTTPX, and Requests. It reports when no spans were produced and preserves the agent's exit code.

Existing spanhound collect, JSONL, OTLP JSON, MCP, and desktop bundle flows remain available. Automatic instrumentation observes supported libraries; arbitrary local functions and approval decisions still require matching instrumentation.

SpanHound v0.5.0 — local OTLP capture

SpanHound can now capture OTLP/HTTP traces locally from agents that already export OpenTelemetry traces.

• Run uvx spanhound collect --trace-dir ./traces to receive JSON or protobuf traces on 127.0.0.1:4318/v1/traces.
• Point an OTLP/HTTP trace exporter at that endpoint, then use spanhound --trace-dir ./traces --check or connect the MCP server to the same folder.
• The receiver accepts gzip, validates requests, limits request size to 8 MiB, and writes private local files.
• The guide now explains how to install the downloaded .mcpb file in a desktop client.

The receiver does not automatically instrument an agent. Existing JSONL and saved OTLP JSON workflows remain supported.

SpanHound v0.4.2

Faster first run

Run uvx spanhound demo to inspect a bundled, recorded shell-boundary failure immediately. The demo shows the tool diagnostic and approval evidence without a clone, trace file, or MCP client.

For your own traces, use uvx spanhound --trace-dir /path/to/traces --check, then connect the MCP server to your client.

This release also adds a UV-backed MCP Bundle asset for desktop clients that support .mcpb. The client asks you to choose a local trace folder during installation.

SpanHound remains a read-only local reader of JSONL spans and saved OTLP JSON exports.

SpanHound v0.4.1 — Registry metadata fix

SpanHound 0.4.1

This patch shortens the MCP Registry description to satisfy its 100-character limit and validates Registry metadata before uploading future PyPI releases. SpanHound 0.4.0 had already published to PyPI; this release publishes the matching package and the new io.github.abhishekash/spanhound Registry entry.

SpanHound reads local JSONL spans and saved OTLP/JSON exports. Run uvx spanhound --trace-dir /path/to/traces --check to verify your trace directory.

SpanHound v0.4.0 — OTLP JSON support

SpanHound 0.4.0

SpanHound is the new name for mcp-trace-inspector. It remains a local, read-only MCP server for inspecting agent runs. Existing Python import paths and command aliases are retained in the new package.

New

• Read saved OTLP/JSON trace exports containing resourceSpans, alongside the existing JSONL span format.
• Use OpenTelemetry GenAI model, token, and tool attributes in run summaries and tool statistics.
• Report OTLP status errors and exception events in failure queries.
• Updated quickstart, sample OTLP export, Docker entry point, and demo media.

Run uvx spanhound --trace-dir /path/to/traces --check to validate a local directory before connecting an MCP client.

v0.3.1 — trace query fixes and first-run check

What changed

• Reject ambiguous trace ID prefixes and show full IDs in run listings.
• Read trace files being written without failing on an incomplete final JSONL line.
• Report unknown model costs as unknown, include all failed spans in run comparisons, and preserve the latest failure diagnostic.
• Add --check to validate a trace directory before connecting an MCP client.
• Clarify the local JSONL format, first-run steps, and the project's scope.

The first PyPI release, 0.3.0, was built from an earlier tag and does not include these fixes. This release makes the published package match the current documentation and tested source.

mcp-trace-inspector v0.3.0

Release source and distributions are built and tested. PyPI publication is pending registration of the mcp-trace-inspector pending Trusted Publisher; the release workflow upload currently fails with PyPI HTTP 400. After the publisher is configured, rerun the failed publish workflow. Existing commands and Python import aliases remain available for compatibility.

AgentHawk v0.2.2

Metadata release for the repository rename to mcp-trace-inspector. The AgentHawk package, CLI, and MCP title remain unchanged; PyPI project links and MCP Registry metadata now reference the renamed repository.

AgentHawk v0.2.1

Full Changelog: v0.2.0...v0.2.1