session-recap

session-recap

English | 中文

[session-recap icon]

Remember what every AI coding session was for.

_{intent · key decisions · what changed · open threads · where to resume, for sessions of any length, from any agent, through MCP, a local viewer or a plugin}

[tests] [release] [npm] [license]

[Claude Code]  [Codex]  [dsh]  [OpenCode]

Projects grow, and so does the number of sessions in each one, often across several agents. After a few weeks a session is a 100 MB transcript you no longer remember: why you started it, what you decided, what was left half done.

session-recap reads a session's whole transcript and writes a short, structured recap. Use it from any MCP client (Codex, Cursor, Zed, Cline, Gemini CLI, OpenCode, Claude Code…), in a local viewer for every agent's sessions, from the command line, or through native plugins for Claude Code, dsh and Codex.

[A recap: an Intent card, a Where to resume card, and tabs for decisions, changes and open threads]

Every recap also has a Map: the phases the work went through (pivots marked) down to "you are here", the key decisions with what was turned down, and lists of what shipped, what is open and what was dropped. Recaps are written in the language you mostly write in.

Quick start

npm install -g session-recapsession-recap mcp-config            # prints the MCP entry for your clientsession-recap serve                 # opens the local viewer

Then, in your agent: "Recap this session", "What was my last session in this project about?", "总结一下这个会话".

No API key needed. Through MCP, the agent you are talking to writes the recap with its own model; session-recap only reads and condenses the transcript. A model of session-recap's own is optional (for the CLI and the viewer's Write recap button).

Or run it without installing: npx session-recap …. session-recap has no dependencies and runs no install scripts. The development version: npm install -g github:BruceXcluding/session-recap --allow-git=all (npm 12 blocks installs from GitHub without the flag).

Requirements: Node.js 20+, python3 (macOS ships one), and for dsh sessions the zstd command or Python 3.14+.

Ways to use it

MCP server

session-recap mcp-config <client> prints the entry for claude, codex, cursor, gemini (and Qwen Code), opencode, zed, vscode or cline, with absolute paths so an IDE started from the dock finds it. For example:

session-recap mcp-config codex >> ~/.codex/config.tomlclaude mcp add --scope user session-recap -- session-recap mcp

Tools:

recap_session also takes format: standup, handoff or pr instead of the recap.

Local viewer

session-recap serve opens http://127.0.0.1:4711 with every agent's sessions by project folder, grouped by time, filtered by agent; each session's Recap and Map; and, with a model configured, Write recap and Update recap. It reads the recaps written by the Claude Code plugin, the CLI and MCP clients alike. It answers only on loopback, to the browser tab it opened.

Claude Code plugin

/session-recap recaps the session you are in; /session-recap browse lists projects and sessions; /session-recap <name> opens a project or a session by name. The pane shows Recap and Map, uses the session's own model (or one picked under Model), and shows your Claude Code sessions first, with All agents one tap away.

Command line

The CLI's own model: SESSION_RECAP_API (anthropic or openai), SESSION_RECAP_BASE_URL, SESSION_RECAP_API_KEY, SESSION_RECAP_MODEL, or ANTHROPIC_*, DEEPSEEK_API_KEY, OPENAI_API_KEY.

Supported agents

Want another? Adding an agent is one reader and one test, or ask for it.

Privacy

• Local. Sessions are read on your machine. Text leaves it only in the model call a recap needs: your agent's (MCP, plugins) or the one you configured (CLI, viewer). No telemetry.
• Secrets removed first. Keys, tokens, private keys, credentials in URLs and PASSWORD=…-style assignments are replaced with [REDACTED:<kind>] before any text reaches a model. --dry-run shows how many. It is pattern matching, so it cannot know every format: see SECURITY.md.
• What is sent is a condensed transcript: your prompts, each turn's final reply, commits and PRs, the files edited. Tool output and intermediate replies are left out.

Token use

Listing never calls a model. On a real 179 MB Claude Code session (220 prompts), a first recap sends about 195k characters (~96k tokens); when the session continues, an update reads only the new prompts. Long sessions are condensed in parts of ≤75k characters, each part's notes cached, and a cheaper model can condense the parts (partModel).

How it works

saved sessions (Claude Code · Codex · dsh · OpenCode)        │hooks/digest.py      one reader per agent → turns (prompt, actions, final reply, edits) → secrets removed        │core/recap.mjs       parts → notes (parallel, cached) → one call writes the recap and its map, or updates it        │MCP server · viewer · CLI · Claude Code pane · dsh tool and /recap      (one cache: ~/.cache/session-recap)

Output is versioned: see docs/data-format.md and schema/recap-v1.schema.json to build your own viewer or exporter.

Contributing

Agents, recap formats, translations and viewers are all welcome. Start with CONTRIBUTING.md; adding an agent and adding a format (one entry in core/formats.mjs) are the easiest first contributions.

License

MIT