meshai-claude-code
EU AI Act Article 12 requires durable records of AI system activity, not best-effort logs. Best-effort telemetry loses events exactly when it matters most: process crashes, OOM kills, network outages. Themeshai-claude-code connector fsyncs every hook event to a local
write-ahead log before anything else happens, so a crash cannot lose an
acknowledged event. Only disk failure can.
Install
Plain
pip install inside a virtual environment also works. On a system
Python, bare pip install fails with error: externally-managed-environment
(PEP 668) on modern Debian, Ubuntu, and Homebrew. Use pipx, a venv, or
pip install --break-system-packages if you understand the tradeoff.Setup
1
Log in
2
Install hooks
~/.claude/settings.json, with a
backup of your existing settings and a clean uninstall path.3
Check status
The daemon starts automatically on your next Claude Code session
(
auto_start_daemon: true by default), or you can run it directly with
meshai-cc-daemon.Architecture
- Hooks run on SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, PreCompact, and Stop. Each hook appends a CRC-framed record to the WAL with a real fsync and exits.
- The WAL (write-ahead log) is local, durable state. Hooks own writes; the daemon only reads.
- The daemon converts WAL events to OpenTelemetry spans and exports them to MeshAI over OTLP. Delivery is at-least-once; span IDs are minted once at hook time, so MeshAI’s ingest dedup makes accounting exactly-once.
Content filtering
Tool content (tool_input / tool_output) is dropped by default. Structural
metadata (event type, tool name, timing, token counts) always flows. Opt in
per tool in ~/.config/meshai/filters.yaml:
Compliance mode
fail_closed: true, a tool call that cannot be durably recorded does
not run: no evidence means no action.
Platforms
macOS, Linux, and WSL (v1). On WSL, state must live on the Linux filesystem, not/mnt/c; this is the default
(~/.local/state/meshai-cc). Native Windows support is planned for v2.