Watch
The local tracehouse agent follows new Claude Code JSONL sessions as they are written.
Claude Code observability, without the plumbing
Claude Code Monitoring & Session Replay
tracehouse records what happened inside every Claude Code session—not only the final answer. Replay prompts, tool calls, tool outputs, subagents, timing and errors in one readable trace.
Automatic recording · local redaction · share only when you choose
live trace · refactor-auth · claude-code
Shareuseragentthinktoolresult
A live product trace. Click a row for the hidden test drive.
How Claude Code monitoring works
From local transcript to a trace you can actually use
Structure
Prompts, reasoning, tool calls and outputs become linked spans with timestamps and status.
Replay
Open the waterfall, search the conversation, share a trace or compare it with another run.
What tracehouse records
The evidence behind the outcome
Prompts and responses
See the user request and every assistant turn in order.
Tool calls
Inspect Read, Edit, Bash, MCP and other tool inputs.
Tool outputs
Keep the result next to the call that produced it.
Subagents
Follow nested work and sidechains without losing the parent run.
Timing
See when each event happened and where the session spent time.
Errors
Surface failed commands, suspicious loops and error propagation.
How to set it up
One install, then keep coding
The installer detects Claude Code, installs the tracehouse binary and configures the background agent. It does not require wrapping each prompt or instrumenting your code.
Read installation details$ curl -fsSL https://tracehouse.ai/install.sh | bashClaude Code transcript directory found
Local secret and PII redaction enabled
New sessions will sync automatically
Session replay
Read the run as a conversation or a waterfall
Move from the original prompt through reasoning, tool calls and results without opening raw JSONL. Timing and nesting keep long sessions understandable, while filters help you isolate the moment a run went off course.
Tool call tracing
See the command, its context and its result
A failed Bash command is useful only when you can see why it ran and what happened next. tracehouse pairs tool inputs and outputs and keeps them in the surrounding Claude Code trace.
Sharing traces and datasets
Send the evidence, not a screen recording
Publish a selected trace as a link for a teammate, reviewer or bug report. You can also share a read-only dataset view when a single run is not enough. Recipients can inspect the replay or row set without needing a tracehouse account. See a real example in this public 160-session agent dataset. Browse additional Fable 5, Claude Code and coding-agent sources in the trace dataset index.
Comparison of runs
Diff good and bad agent runs side by side
Compare two attempts to see which prompts, tools, timings and outcomes changed. It is a faster path from “this run felt worse” to a concrete, reviewable difference.
tracehouse vs raw OpenTelemetry setup
Choose the layer you need
Claude Code can export telemetry through OpenTelemetry. That is a good foundation when your team wants to operate its own collector, storage and dashboards. tracehouse is for teams that want an agent-aware product experience immediately.
| Capability | tracehouse | Raw OTel stack |
|---|---|---|
| Collection | Transcript watcher and agent parser included | Configure exporters and collector |
| Session replay | Purpose-built conversation + waterfall | Build or adapt a trace UI |
| Tool inputs and outputs | Paired and readable by default | Model attributes and queries |
| Sharing and comparison | Included product workflows | Implement in your backend/UI |
| Infrastructure control | Managed application | Fully self-operated |
FAQ
Claude Code monitoring questions
What is Claude Code monitoring?
Claude Code monitoring records the events inside a coding session—prompts, assistant messages, tool calls, results, timing and errors—so you can inspect the run after it finishes instead of relying on the final diff alone.
Does tracehouse require me to change how I use Claude Code?
No. The tracehouse agent watches the JSONL transcripts Claude Code already writes. After one-time setup, normal Claude Code sessions are recorded automatically in the background.
Can I monitor tool usage and tool outputs?
Yes. Tool calls are displayed with their inputs, outputs, status and position in the surrounding conversation, including Bash, file operations and MCP tools captured by the transcript.
Can teammates view a trace without an account?
A trace can be shared with a public link. Anyone with the link can inspect the replay without signing in; you decide which trace to publish.
How is this different from a raw OpenTelemetry setup?
OpenTelemetry is a flexible transport and data model. tracehouse is a ready-to-use product for coding-agent sessions: transcript ingestion, agent-aware parsing, replay UI, sharing, comparisons and analysis are included.
Know what the agent did before you ship it
Install tracehouse, run Claude Code normally, and open the complete session replay when you need it.