Forty-three coding agents — including Claude Code, Codex, Cursor, OpenClaw, Gemini CLI, opencode, Cline, Copilot CLI, Roo Code, Kilo Code, aider, Goose, Qwen Code, Kimi Code, Antigravity, Grok Build, pi, omp, DeepSeek Harness, Hermes, Kiro, Reasonix and Zed — already write every session to local files. deja starts full: it indexes months of history from before you installed it and serves relevant context back over MCP and hooks, with no model and no capture step.

One Go binary. No LLM, no embeddings, no API key. The network is used only by commands you run for it, such as deja update and deja sync ssh, and by a daily release check, which DEJA_OFFLINE=1 turns off. 58% fewer tokens on a task this machine had already solved (53,558 against 126,222 with nothing wired, and 52,815 against 103,443 on a later build with the arms run alternately, eleven runs an arm each time). 88.9% hit@1 on LongMemEval-S (470-question cleaned set), millisecond lookups over gigabytes of history.

npx @vshulcz/deja-vu "connection pool exhausted"   # search, no install
npm install -g @vshulcz/deja-vu                    # then: deja install --auto

deja install --auto wires MCP recall into every coding agent it finds on the machine and turns on session-start recall where the agent supports it. Ten seconds to install, about ten to index, and the next session already knows. Use --all for the MCP tools without the session-start hook. deja on its own opens a search screen over every agent's history, where you can read a session, resume it, or continue it in another agent.

Or wire the stdio MCP server directly in any MCP client (mcp.json / claude_desktop_config.json):

{
  "mcpServers": {
    "deja": {
      "command": "npx",
      "args": ["-y", "@vshulcz/deja-vu", "mcp"]
    }
  }
}

Full documentation, the harness matrix and the benchmarks: github.com/vshulcz/deja-vu · vshulcz.github.io/deja-vu

MIT

Found it useful? Star deja-vu on GitHub.