SpanHound README

SpanHound

Capture local agent traces. Ask your MCP client what happened. SpanHound (formerly mcp-trace-inspector) captures supported Python agent calls, receives OpenTelemetry over local OTLP/HTTP, and exposes read-only MCP tools for failures, approval decisions, token usage, and run comparisons. It reads JSONL spans and OTLP JSON exports as described in the user guide, including traces from agent-harness.

Demo: Watch the full 48-second video · Read the reproducible story. This is a scripted regression trace, not a live-model benchmark.

The project's concrete origin was a shell-boundary regression: the harness blocked cat ../outside.txt, but the old trace recorded only that run_shell was called, not the returned diagnostic. SpanHound and the harness now preserve and query that bounded error.

Try it in one command

bash

uvx spanhound demo

text

SpanHound demo — recorded agent failure
Tool: run_shell
Failure: error: shell path escapes workspace root: '../outside.txt'
Approval: approve (auto)
Trace: 563e779ee8fa2d7069fc63ed2cf9c8ec (6 spans)

The demo shows a recorded run_shell failure and its approval evidence from a bundled scripted regression trace. It needs no clone, account, trace backend, or MCP client. To inspect your own traces, point SpanHound at a directory of JSONL spans or saved OTLP JSON exports:

bash

uvx spanhound --trace-dir /path/to/traces --check

--check prints the runs it found and exits. Connect an MCP client with the configuration below to ask questions about those runs.

Explore the repository fixtures

bash

git clone https://github.com/abhishekash/spanhound.git
cd spanhound
uvx --from . spanhound --trace-dir examples --check

--check also reports a missing directory or malformed complete JSONL or OTLP JSON record.

To try an OTLP JSON export, run the same command with --trace-dir tests/fixtures. The sample export includes a model call, token usage, a tool call, and an exception.

Capture your own traces

For a Python agent, install SpanHound's optional auto instrumentation in the same environment as the agent, then run it through SpanHound:

bash

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

The command starts and stops a local receiver, configures OTLP export, and captures supported library calls without changes to agent.py. The optional package includes instrumentation for OpenAI, Anthropic, OpenAI Agents, HTTPX, and Requests. It reports when no spans were produced. Your Python agent and SpanHound must share one environment; for a uv project, use uv add 'spanhound[auto]' and uv run spanhound run -- python agent.py. See the Python capture guide.

For other runtimes, or agents that already export OpenTelemetry traces, start the receiver in one terminal:

bash

uvx spanhound collect --trace-dir ./traces

Point an agent that already uses an OpenTelemetry trace exporter at the receiver, then run the agent:

server.ts

export OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=http://127.0.0.1:4318/v1/traces
export OTEL_EXPORTER_OTLP_TRACES_PROTOCOL=http/protobuf
# Start your instrumented agent here.

SpanHound stores each accepted OTLP/HTTP trace request as a local JSON file. Run uvx spanhound --trace-dir ./traces --check to see the captured runs, then connect your MCP client to that folder. The receiver accepts OTLP/HTTP JSON or protobuf on loopback; it does not add instrumentation to an agent. See the capture guide for limits and privacy.

Where it fits

For a live terminal and browser view of local Codex/Claude Code sessions and received OTLP telemetry, try Agent Observatory. SpanHound provides trace investigation through MCP using JSONL and OTLP JSON files.

Use this when you want a local trace view through MCP without setting up a trace backend. Capture supported Python library calls in one command, receive OTLP/HTTP from another runtime, or read the supported JSONL or OTLP JSON formats. Focused queries include recorded approval decisions, bounded failure diagnostics, and run comparisons.

MCP Agent Trace Inspector records steps into a local SQLite database and provides a dashboard and exports. SpanHound auto-instruments supported Python libraries or receives OTLP/HTTP traces, then queries the captured files through MCP. It does not automatically trace arbitrary agent functions or approvals.

For traces stored in Jaeger, Tempo, or Traceloop, use a backend-connected tool such as OpenTelemetry MCP Server. For a trace UI, instrumentation, and evaluations, see Phoenix. SpanHound does not read backend-specific JSON exports.

AI-native use cases

Question during an agent run	MCP tool	Evidence returned
"Why did yesterday's run stall?"	`list_runs` → `slowest_spans`	Run IDs and the longest model or tool spans.
"Did the agent act after I denied the write?"	`approval_log` → `span_tree`	The recorded decision and subsequent execution path.
"How many model tokens did this run use?"	`token_usage` → `span_tree`	Run-level usage totals and the span tree for context.
"Which tool is failing or timing out?"	`failure_report` / `tool_stats`	Redacted diagnostics, error rates, denials, and p95 latency.
"Who/what was allowed to act?"	`security_audit`	Caller, tenant, server provenance, risk, and missing-evidence findings.
"What is happening in the run right now?"	`recent_activity`	Cursor-based polling of newly written spans.
"Did this model/version regress?"	`compare_runs`	Cost, latency, tokens, tool calls, and failure deltas.

The tools read stored traces and can poll files being written, but they do not push events or infer intent from the final answer. Start with the installation and user guide, then see the real trace fixture and the tool descriptions for the query contract.

Install & run

bash

uvx spanhound --trace-dir ./traces
# or, for local development:
git clone https://github.com/abhishekash/spanhound
cd spanhound
uvx --from . spanhound --trace-dir ./traces

For a desktop client that accepts MCP Bundles, download the .mcpb asset from the latest release and install the downloaded file in the client; the download link does not open a web page. In Claude Desktop, use Settings → Extensions → Advanced settings → Install Extension… and select the .mcpb file. The bundle asks you to choose a local trace folder. It uses UV to install its Python dependencies on first run. SpanHound reads the folder locally; the connected client receives the evidence returned by its MCP tools.

If you previously ran uvx mcp-trace-inspector, change that command to uvx spanhound to receive this release. The old PyPI distribution stays at 0.3.1; the new package retains its Python import paths and command aliases for clients that install spanhound explicitly.

Client configuration

Claude Desktop (claude_desktop_config.json):

config.json

{
  "mcpServers": {
    "agent-traces": {
      "command": "uvx",
      "args": ["spanhound", "--trace-dir", "/path/to/traces"]
    }
  }
}

pi (~/.pi/agent/settings.json):

config.json

{
  "mcpServers": {
    "agent-traces": {
      "command": "uvx",
      "args": ["spanhound", "--trace-dir", "/path/to/traces"]
    }
  }
}

agent-harness (mounted as gated tools):

bash

harness run "Why was my last run slow?" --mcp "uvx spanhound --trace-dir ./traces"

Tools

Tool	Use it when
`list_runs`	Starting out — recent runs with task, model, duration, cost, decision counts
`run_summary`	One run at a glance (accepts trace-id prefix)
`span_tree`	"What did the agent actually do?" — nested shape of the run
`slowest_spans`	"Why was it slow?" — top-k spans by duration
`approval_log`	HITL audit — every approve/deny/edit, who decided, and the rationale
`token_usage`	Cost questions — aggregated across runs or per-run
`search_spans`	Find spans by tool name, file path, "denied", …
`failure_report`	Find actionable, bounded, redacted errors and failed tool calls.
`tool_stats`	Rank tools by volume, error rate, denials, and latency.
`security_audit`	Audit caller identity, tenant, server provenance, risk, and approval evidence.
`recent_activity`	Poll new spans with a cursor while a run is active.
`compare_runs`	Compare selected traces for regressions across models or versions.

Tool descriptions are written as prompts (when-to-use, not just what-it-does) — descriptions are the interface for agent-called tools.

Example session (real fixture trace)

Code

> list_runs
[{ "trace_id": "f920798dd2558fcb84f9977736432a8b", "task": "Summarize the workspace's notes…",
   "tool_calls": 4, "human_decisions": 2, "stopped_reason": "completed" }]

> approval_log
[{ "tool": "write_file", "decision": "approve", "approver": "auto", … },
 { "tool": "run_shell",  "decision": "approve", "approver": "auto", … }]

Design

Code

traces/*.jsonl ──┐
traces/*.json  ──┴─▶ mcp_trace.core (local query functions)
                          │
                   mcp_trace.server (MCP adapter)
                          │
                    stdio (JSON-RPC)

core/server split: all logic is pure functions over parsed spans; the MCP layer only parses args and JSON-encodes results. Tests hit both layers.
trace_id prefixes: agents fumble full 32-char hex ids; every tool accepts prefixes.
bounded output: diagnostics are truncated and obvious credentials are redacted before query results leave the server.
honest cost totals: token_usage returns cost_known: false and a null cost when any model span lacks a cost or marks pricing as unavailable; unknown pricing is never presented as $0.
cursor polling: recent_activity makes the snapshot reader useful while a run is still writing spans.
The demo fixture (examples/example_trace.jsonl) is a real agent-harness run, not hand-written.

Honest limitations

MCP stdio transport only (Streamable HTTP plus authenticated remote access is the next transport boundary)
live polling re-reads snapshots; it is not a push subscription
security audit can only report identity/provenance that the trace producer records
read-only tools; trace mutation (annotations) is roadmap
The Python query and server modules are available as mcp_trace; older import and command aliases remain for existing clients.

First-run feedback

Tell us what happened when you tried it: your MCP client or trace producer, the question you were trying to answer, and any missing evidence. Share a redacted excerpt when a file fails to load.

License

MIT