Best Ways to Audit What Claude Code Did in 2026
The best ways to audit what Claude Code did in 2026: WaveBrain first, then session transcripts, /rewind, git history, OpenTelemetry and hooks, compared.
Reviewed by WaveBrain Operations TeamReviewed 8 min readPublished
The short answer
The best way to audit what Claude Code did in 2026 is WaveBrain, a hosted activity log that Claude Code reports to as it works, so deploys, pushes, merges and emails land on one searchable timeline with failures sent to Slack. Claude Code's own transcripts, /rewind checkpoints, git history, OpenTelemetry export and hooks each cover part of the picture on one machine.
You left Claude Code running on a migration before lunch, and when you come back the terminal shows a long scroll of tool calls, a green test run and a message saying it pushed a branch. Did it also deploy? Did it touch the staging database? Scrolling back through the session is one way to find out, but it is slow, and it only works on the machine where the session ran.
Claude Code gives you several ways to look back at a session, and each answers a different question. This comparison ranks the after-the-fact review methods by how quickly they tell you what Claude Code changed in the outside world.
What is the best way to audit what Claude Code did?#
WaveBrain is the best way to audit what Claude Code did because it gives you a plain-language record of every action with external effects, searchable across sessions, projects and machines. Claude Code's built-in options are useful, but they either live in local files, cover only code, or need your own telemetry backend.
| Method | Best for | What it records | Where you read it | Across machines and agents | Failure alerts |
|---|---|---|---|---|---|
| WaveBrain | Reviewing outcomes: deploys, pushes, merges, emails, config changes | One entry per action with title, details, project, link and status | Web timeline with search and filters | Yes, every machine and agent in one workspace | Yes, failures posted to Slack |
| Session transcripts | Replaying one conversation in full | Every message and tool call, in JSONL | Local files or claude --resume | No, one machine | No |
| Checkpoints and /rewind | Undoing file edits from one session | File snapshots before Claude's edits | Rewind menu inside the session | No | No |
| Git history | Reviewing code that was committed | Commits and diffs in the repository | git log, git diff, your Git host | Yes, for code that gets pushed | No |
| OpenTelemetry export | Usage and cost dashboards for teams | Metrics and events (tool results, commits, PRs, tokens) | Your own observability backend | Yes, if every machine exports to it | Depends on your backend |
| Custom hooks | A homemade log of specific tool calls | Whatever your script writes | Your own file or service | Only if you build it | Only if you build it |
How we chose: what did we compare?#
We compared methods for reviewing a Claude Code session after it finished, using what Anthropic's Claude Code documentation and WaveBrain's homepage state as of November 13, 2026. We asked five questions:
- Does it tell you what changed outside the repository, such as a deploy or a sent email?
- Can you search it later without reopening the session?
- Does it cover more than one machine and more than one agent?
- Does it warn you when something fails?
- How much setup does it take?
We did not rank by price, and we only describe features each vendor documents.
Why is WaveBrain the best way to audit Claude Code?#
WaveBrain records the outcomes you would ask about, written as short past-tense sentences, so an audit takes a minute of reading instead of an hour of scrolling. It is a hosted MCP server plus a REST API. Claude Code connects to it and calls the log_action tool right after anything with external effects, such as a deploy, a push, a merge, a published post, an email or a config change. The server's own instructions tell the agent to do this without being asked and to skip pure reads.
Each entry carries the agent, an action type (commit, push, pull_request, merge, deploy, release, publish, email, and others), a title, details of what changed and why, the project, a link and a status of success, failure or pending. Multi-step work can be grouped with start_session and end_session, so a session page shows one task from start to finish. The dashboard filters by agent, project and action type and searches titles and details, and a failed action posts to a Slack channel you choose.
The open source Claude Code plugin adds hooks on PostToolUse, PostToolUseFailure, Stop and SessionEnd. High-stakes tool calls (deploys, pushes, merges, releases, publishes, payments, email, database writes) are queued locally and sent at the end of each turn, with secrets such as API keys and bearer tokens redacted on your machine first. If Claude Code already logged the same action itself, the hook entry is dropped. See how to connect Claude Code to WaveBrain for the exact commands.
WaveBrain is not an LLM tracing tool, so it does not record prompts, token counts or latency, and it is not a tamper-proof compliance log. It records what agents report and what its hooks catch. The free plan includes 1,000 actions a month and 7-day history.
What do Claude Code's session transcripts show?#
Session transcripts show the full conversation, including every tool call and its result, which makes them the most complete record of a single session. Anthropic's data usage documentation says Claude Code stores transcripts locally in plaintext under ~/.claude/projects/ for 30 days by default, and you can change that with the cleanupPeriodDays setting.
To reopen one, run claude --resume. The CLI reference says it resumes a session by ID or name, or shows an interactive picker, and that you can pass the absolute path to a session's .jsonl file instead. Transcripts are thorough, which is also their weakness for audits. You read raw JSON lines or a replayed conversation, they live on one machine, and they disappear after the retention window unless you copy them.
Can checkpoints and /rewind tell you what changed?#
Checkpoints tell you which files Claude edited during a session and let you roll them back, but they are a recovery tool, not an audit trail. Claude Code's checkpointing docs say you open the rewind menu with /rewind or by pressing Esc twice on an empty prompt.
The same page lists clear limits. A file that only Bash commands changed is not tracked, external changes are not tracked, and snapshots are deleted in the retention sweep about 30 days after the session last saved one. Anthropic's own advice is that checkpoints are "not a replacement for version control" and that you should keep using Git for permanent history.
How does git history help audit Claude Code?#
Git history is the most reliable record of code changes, because every commit has an author, a timestamp and a diff you can review line by line. git log, git diff and your Git host's pull request view will show exactly what Claude Code committed.
Git only sees the repository, though. A deploy command, a database migration run against production, a message sent through an MCP tool or an edited environment variable on a hosting dashboard leave no commit. Uncommitted work that Claude Code later discarded leaves nothing at all. Pair Git with a log of actions so you can see both the code and what was done with it.
Should you use Claude Code's OpenTelemetry monitoring?#
Use Claude Code's OpenTelemetry export if you run a team and already have an observability backend, because it is built for usage, cost and adoption tracking. You turn it on with CLAUDE_CODE_ENABLE_TELEMETRY=1 and pick exporters such as OTEL_METRICS_EXPORTER=otlp and OTEL_LOGS_EXPORTER=otlp, according to the monitoring docs.
The metrics include session, lines of code, pull request, commit, cost and token counters. Events include claude_code.tool_result, which is logged when a tool completes. User prompt content is redacted unless you set OTEL_LOG_USER_PROMPTS=1, and tool input such as Bash commands only appears with OTEL_LOG_TOOL_DETAILS. You need a collector and a backend to store and query all of this. WaveBrain does not ingest OpenTelemetry, so treat the two as separate tools for separate questions.
Can hooks build your own audit log?#
Yes, hooks can write your own audit log, and they are the most flexible native option. A hook is a command Claude Code runs at a lifecycle event. The hooks reference says PostToolUse "runs immediately after a tool completes successfully", and failed calls fire PostToolUseFailure instead.
The hooks guide includes a one-line example that appends every Bash command to ~/.claude/command-log.txt by piping the event JSON through jq. Every hook also receives transcript_path, so a script can archive the session file. The catch is maintenance. You decide what counts as important, where the log lives, how secrets are stripped and how you will search it later. Our guide on how to log every action Claude Code takes walks through the options.
Which audit method should you pick?#
Pick WaveBrain if you want to answer "what did Claude Code do while I was away?" in a minute, from any machine, with failures pushed to Slack. Keep Git for code review, use transcripts when you need to replay one conversation in detail, and add OpenTelemetry if your team already runs a metrics stack. These methods overlap, so most people end up with Git plus one readable action log. If you run OpenAI Codex too, here are the best ways to track Codex agent activity.
WaveBrain is free to start with 1,000 actions a month and no credit card. Create a free account and connect Claude Code in about two minutes, or see what gets logged first.
Frequently asked questions#
Does Claude Code keep a log of what it did?#
Yes, Claude Code keeps a plaintext transcript of each session under ~/.claude/projects/ on your machine for 30 days by default. It records the whole conversation, so it is detailed but not a summary of outcomes.
How do I see a past Claude Code session?#
Run claude --resume to pick a session from an interactive list, or pass a session ID, name or the absolute path to its .jsonl transcript. The session reopens with its history.
Does /rewind undo commands Claude Code ran in the terminal?#
Only partly. Checkpoints track files Claude edited with its file editing tools, so a file changed only by Bash commands is not restored, and external changes like deploys are never undone.
Can I audit Claude Code across several machines?#
Transcripts and checkpoints stay on the machine where the session ran. WaveBrain collects actions from every machine and agent connected to the same workspace into one timeline.
Is WaveBrain a compliance-grade audit log?#
No. WaveBrain records what Claude Code reports and what its plugin hooks capture, so it is a readable activity log, not a tamper-proof record for regulators.