v0.4.2 — MCP-registry launch fix (uvx claude-replay mcp)

Registry/packaging fix so uvx — and MCP directories (e.g. Glama) that auto-launch via uvx — can install and run the stdio server. No functional change.

• Registry launch config: the mcp subcommand moved from runtimeArguments to packageArguments → uvx claude-replay mcp (was the invalid uvx mcp claude-replay). Unblocks directory auto-installers that reported "cannot be installed."
• Pinned hatchling<1.30 so the OIDC build emits uploadable Metadata-2.4.

v0.4.1 — installable again (mcp<2) + dependency hardening

Dependency-hardening patch. No feature or API changes.

The important fix

mcp was declared mcp>=1.0.0 with no upper bound. Once mcp 2.0 shipped — removing the low-level @server.list_tools() / @server.call_tool() decorator API the server is built on — a fresh pip install claude-replay (or uvx claude-replay mcp) resolved mcp 2.0 and crashed on import, so serve and mcp were dead on 0.4.0. 0.4.1 caps mcp<2. If you're on 0.4.0, upgrade.

Also

• Upper-bounded every other runtime, TUI, and dev dependency below its next major.
• Pinned the ruff lint rule set ([tool.ruff.lint] select = E4/E7/E9/F) so an unpinned ruff can't fail ruff check . on unchanged code.
• Fixed two time-based prune tests that hardcoded a "recent" date and would rot as real time passed.

374 tests pass in a clean venv. Full changelog: https://github.com/constripacity/Claude-Replay/blob/main/CHANGELOG.md

v0.4.0 — Reliability & cross-session insight

First release recorded against real Claude Code sessions (not synthetic data).

Added

• claude-replay doctor — a self-check that answers is Replay actually recording? Verifies hooks are installed, the claude-replay command resolves on PATH (the #1 silent-failure mode — a hook Claude Code can't find records nothing), the DB exists, and sessions are being recorded. Plus an install-time PATH warning.
• Cross-session analytics — claude-replay stats, a replay_stats MCP tool (ten tools now), and /api/stats: total/avg tool calls, overall error rate, the death-cause breakdown ("why your sessions end"), tool mix, per-project rollups, and a day-by-day trend.

374 tests, ruff clean. Pre-release audit clean (secrets, deps, wheel/sdist contents, twine).

Install: pip install claude-replay · then claude-replay doctor

Claude Replay 0.3.0

Claude Replay 0.2.0

v0.1.0 — Claude Replay

First public release of Claude Replay — a passive session checkpoint and recovery layer for Claude Code. Hooks record every session to a local SQLite store, and a CLI, five MCP tools, a web dashboard, and a terminal UI read it back. Offline-first; nothing leaves your machine.

Install

pip install claude-replay

With the terminal UI:

pip install "claude-replay[tui]"

Then wire it into Claude Code and start the server:

claude-replay install     # merges the hooks into ~/.claude/settings.json (idempotent)
claude-replay serve       # dashboard + MCP tools on http://127.0.0.1:8766

What's included

• Store — SQLite-backed sessions, events, and checkpoints, with per-session monotonic sequencing and 8 KB-per-event payload truncation. DB at ~/.claude-replay/sessions.db (overridable via CLAUDE_REPLAY_DB).
• Hooks — PreToolUse / PostToolUse / Stop handlers that record tool calls, auto-checkpoint every 10 tool results, and write a final checkpoint on stop. Completes in well under 50 ms and never breaks the agent it records.
• Recovery — generate a paste-ready resume brief (objective, work done, pending step, files touched) from any session.
• Export — render a session as a self-contained HTML trace.
• Server — Starlette app on port 8766: five replay_* MCP tools over SSE, a JSON API, and a static web dashboard.
• Terminal UI — a Textual session browser over the JSON API, with one-key resume and export.
• CLI — install, uninstall, status, sessions, resume, export, serve, tui, reset.
• MCP tools — replay_status, replay_checkpoint, replay_resume, replay_sessions, replay_export.

Links

• PyPI: https://pypi.org/project/claude-replay/0.1.0/
• Full changelog: CHANGELOG.md