Privacy
Aftertrace is local-first. The Go engine never opens a network connection of its own. The only sockets are the ones you bind for OTLP and the UI. The optional Electron desktop shell (desktop/) may contact GitHub Releases to look for an app update; it does not send usage, traces, or crash reports.
What is redacted by default
At ingest, internal/redact deletes keys from every span, span event, log record, and metric data point, including resource attributes.
Content keys, dropped unless --store-content. The full list is ContentKeys and ContentPrefixes in internal/redact/redact.go:
Claude Code user_prompt prompt prompt_text response response.model_output tool_input tool.output output new_context body
GenAI semconv gen_ai.input.messages gen_ai.output.messages gen_ai.system_instructions gen_ai.tool.definitions
gen_ai.tool.call.arguments gen_ai.tool.call.result gen_ai.retrieval.query.text gen_ai.retrieval.documents
gen_ai.prompt gen_ai.completion gen_ai.request.stop_sequences gen_ai.prompt.* gen_ai.completion.*
Claude SDK anthropic.tool_result.content
OpenLLMetry traceloop.entity.input traceloop.entity.output
OpenInference input.value output.value llm.prompts llm.tools tool_call.function.arguments tool.parameters
llm.input_messages.* llm.output_messages.*
Vercel AI SDK ai.prompt ai.prompt.messages ai.response.text ai.response.toolCalls ai.response.reasoning ai.toolCall.args ai.toolCall.result
Pydantic AI pydantic_ai.all_messages
Gemini CLI request_text response_text function_args
Those names are the ones Claude Code uses for prompt text, responses, tool input, tool output, detailed-beta context, and raw API bodies. See Claude Code monitoring. Local history import never reads those fields into the store at all, even with --store-content: it keeps timestamps, model, tokens, tool names, ids and durations.
Identity keys, always dropped, including with --store-content:
user.email
user.account_uuid
user.account_id
user.id
organization.id
process.owner
user_prompt_length and other non-content attributes are kept. Claude Code still has to send an attribute for Aftertrace to store it. Prompt text is off at the source unless you set OTEL_LOG_USER_PROMPTS=1; see claude-code-setup.md.
OTEL_LOG_TOOL_DETAILS=1 is a separate gate. It sends Bash commands, file paths, and skill / MCP names. Those keys are not in the lists above, so Aftertrace stores them. That is why findings can name Bash(npm test:*).
--store-content
aftertrace serve --store-content and aftertrace import --store-content keep the content keys. Identity keys still go. The serve log line is:
redact=on store_content=false (no telemetry)
store_content=true only after you pass the flag.
Two gates, both default off: Claude Code does not emit content, and Aftertrace deletes the keys if they arrive anyway.
Localhost-only binding
--listen defaults to 127.0.0.1. OTLP gRPC, OTLP HTTP, and the UI all bind that host. Nothing on the LAN can reach the receivers unless you pass --listen 0.0.0.0.
There is no TLS and no auth. Team mode on a shared host is a private network decision, not a Aftertrace feature.
No telemetry, no phone-home
The Go binary does not check for updates, crash-report, or send usage. POST /api/sql is read-only and sandboxed. The one exception is Ask, which is off unless you turn it on. The desktop shell's "Check for updates" (and a silent startup check) talks only to GitHub Releases for the app package; turn it off by not using the desktop app.
Ask
aftertrace serve --ask with ANTHROPIC_API_KEY set turns on plain-English questions on the Query page. Claude writes SQL, Aftertrace runs it in the sandbox below, and Claude answers from the rows. What leaves the machine, to the Anthropic API only:
- The question you typed.
- The schema: table and view columns, and the span and event names present with their counts.
- The rows of each query it runs, at most 50 per query, with every string cut to 200 characters.
The prompt tells Claude not to select prompt, response or tool content, and the 200-character cut keeps long content out even if it does. Nothing is sent when --ask is off, and the Query page says so. The model is claude-opus-5-5 unless AFTERTRACE_ASK_MODEL names another.
SQL sandbox
The Query page's raw SQL box runs your SQL on the store, so it is locked to the store. DuckDB is opened with external access off and its configuration locked: no reading files (read_text('/etc/hosts') fails), no writing files, no ATTACH, no extension INSTALL or LOAD, no network. Aftertrace checks that the statement is a single SELECT and runs it in a transaction that is rolled back. That matters most with --listen 0.0.0.0, where anyone who can reach the UI port can send SQL. Details: api.md.
Where the DB lives
Default file: $HOME/.aftertrace/aftertrace.duckdb. Override with --db. Alongside it DuckDB writes a write-ahead log, *.duckdb.wal.
After ./aftertrace serve --db ./data/aftertrace.duckdb:
./data/aftertrace.duckdb
./data/aftertrace.duckdb.wal
One process at a time. A second serve, demo, or import on the same path fails with a conflicting lock.
How to delete it
Stop aftertrace. Remove the DuckDB file and its WAL.
# default path
rm -f "$HOME/.aftertrace/aftertrace.duckdb" "$HOME/.aftertrace/aftertrace.duckdb.wal"
rmdir "$HOME/.aftertrace" 2>/dev/null || true
# if you used the quickstart path
rm -f ./data/aftertrace.duckdb ./data/aftertrace.duckdb.wal
That is the whole store. There is no remote copy.