Docs
Set up in two minutes
One hosted vault. One command per machine. No daemon, nothing in your agent's request path that can break you.
1. Link a machine
npx aipocalypse login
Prints a code; approve it in the browser. The CLI gets a device-bound key stored at ~/.aipocalypse/config.json (mode 600).
2. Connect your agents
npx aipocalypse install claude # plugin: hooks + MCP + skill, OpenTelemetry (prompts off) npx aipocalypse install codex # hooks.json + [mcp_servers] in config.toml, then /hooks to trust npx aipocalypse install cursor # hooks.json + mcp.json npx aipocalypse pull --targets gemini,copilot # rules, memory snapshot, skills as files
Hook configs on disk only ever carry an ingest key: a leaked laptop config can report events but cannot change your vault.
3. Use it
npx aipocalypse remember "We deploy with railway up --service web" --pin npx aipocalypse push rule style ./STYLE.md --targets claude,codex npx aipocalypse push --skills # every SKILL.md folder, scanned before it ships npx aipocalypse push --memory-import # Claude Code auto-memory → vault npx aipocalypse pull # on any machine; also runs at session start when stale npx aipocalypse graph push # Graphify graph.json → your project constellation
Inside any agent: the aipocalypse MCP server offers memory_add, memory_search, rules_get, vault_pull, status. Claude Code and Codex also get relevant memories injected on every prompt.
Pro
- Phone approvals: Claude Code's permission prompts arrive in Telegram with Allow / Deny. 90 s to answer; otherwise the terminal prompt proceeds as usual. Never on by default.
- GitHub backup: a nightly commit of your vault to a private repo you own, using a fine-grained token scoped to that repo.
- Smart memory: semantic recall from an open-source embedding model that runs inside our server — your notes are never sent to a third-party AI.
- Digest: one morning message with sessions, failures, spend and busiest projects.
- Skill guard (everyone): every session start syncs new or changed local skills through the trust scanner and pings you only if something is wrong.
What we see
- Metadata tier (default): event kinds, tool, model, tokens/cost, tool names, hashed session and project ids. Prompt text is used for memory retrieval and discarded.
- + paths: file paths relative to the project root.
- + prompts: prompt and error text (≤ 8 KB). Never tool inputs/outputs or diffs.
- Memories, rules and skills are stored as you wrote them. Alert channel configs, exports and repo graphs are encrypted at rest.
- Export everything as a zip any time; delete everything in one click. No grace period, no lock-in.
If we are down
Hooks time out in ≤ 5 s and your agent continues. Rules and skills are real files on disk; the CLI applies its last cached pull. Nothing about your setup depends on us being up.