aiyou

https://geniusrise.github.io/you

Export, maintain and analyze all your AI coding-agent chat history — Claude Code, OpenCode, Codex, Charm Crush, Gemini CLI, Pi — in one normalized, git-versioned store on disk. Import your ChatGPT / Claude.ai / Gemini dumps alongside them. Then run an LLM swarm over everything to distill your personality, preferences and workflow into a portable SKILL.md your agents can use to align with you.

Local-first: the store is a plain git repo under ~/.aiyou. No daemon, no server, no network — except when you run aiyou profile.

Install

# Arch / Manjaro (AUR)
yay -S aiyou

# Debian / Ubuntu
sudo dpkg -i aiyou_0.2.0_amd64.deb   # from GitHub releases

# Fedora / RHEL
sudo dnf install aiyou-0.2.0-1.x86_64.rpm

# From source
cargo install --git https://github.com/geniusrise/you
# shell completions: aiyou --shell bash|zsh|fish

Quickstart

aiyou init                # create ~/.aiyou (git repo + config)
aiyou sync                # harvest all local harness chats (incremental, idempotent)
aiyou status              # what's in the store

# import your hosted dumps (re-import newer dumps any time — no duplicates)
aiyou import chatgpt   ~/Downloads/chatgpt-export.zip
aiyou import claude-ai ~/Downloads/claude-export/conversations.json
aiyou import gemini    ~/Downloads/takeout/MyChat.json

# run the swarm: map-reduce over all chats -> profile/traits.json + profile/SKILL.md
export OPENAI_API_KEY=...          # or any OpenAI-compatible provider, see below
aiyou profile

# install the generated skill into a harness
aiyou profile install --target claude      # or opencode | crush | codex, or --agents-dir <path>

Keep it updated (cron)

aiyou sync is incremental and idempotent — safe to run as often as you like. Print ready-made crontab lines:

aiyou cron
# 0 3 * * * AIYOU_HOME=/home/you/.aiyou /usr/local/bin/aiyou sync
# 0 4 * * 0 AIYOU_HOME=/home/you/.aiyou /usr/local/bin/aiyou profile --if-changed

Add a git remote in ~/.aiyou yourself if you want off-machine backups — aiyou never pushes.

What gets stored

~/.aiyou/
  aiyou.toml                     # config
  chats/<source>/<project>/      # per session: <id>.jsonl (canonical) + <id>.md (rendered)
  imports/manifest.json          # every dump ever imported (hash, date, counts)
  profile/
    traits.json                  # canonical machine-readable trait profile
    SKILL.md                     # generated skill for your agents
    runs/<ts>.json               # audit trail of each profile run

• .jsonl = line 1 session header, then one message per line, sorted by timestamp — stable, diff-friendly.
• .md = human-readable render, secrets redacted by default (API keys, tokens). Set redact_all = true to redact the canonical JSONL too.
• Sources are read-only; deleted-from-source sessions are kept in the store (archive semantics).

Supported sources

Harvest (read from disk)	Import (from dumps)
Claude Code, OpenCode, Codex, Charm Crush, Gemini CLI, Pi, Antigravity, Goose, GitHub Copilot (CLI + VS Code chat), Cursor, Amp, Continue, Zed	ChatGPT (conversations.json / .zip), Claude.ai (conversations.json), Gemini (Takeout JSON)

Missing a source (e.g. you don't use Codex)? It's skipped with no errors. Paths are overridable in aiyou.toml.

The profile swarm

aiyou profile shards your chats into batches (~120k chars each) and runs N concurrent workers (configurable) that extract grounded trait candidates (personality, communication, preferences, workflow, tech-stack, values, goals — each with evidence quotes and confidence). A reduce pass merges candidates with your existing traits.json into a canonical profile. Re-runs are incremental: only chats changed since the last run are mapped.

Any OpenAI-compatible endpoint works:

[profile]
provider_base_url = "https://api.openai.com/v1"
api_key_env       = "OPENAI_API_KEY"
model             = "gpt-5.2"
workers           = 4
map_context_chars = 120000
exclude_sources   = []   # e.g. ["chatgpt"] to leave imports out

Point provider_base_url at OpenRouter, a local vLLM, or an Anthropic-compat proxy — anything speaking /chat/completions with response_format: json_object.

Every run writes an audit record (profile/runs/) and a git commit, so profile history is versioned too.

Privacy

• Everything stays local; LLM calls happen only during aiyou profile, only to the endpoint you configure.
• Markdown renders are secret-redacted before they're written (and therefore before they're ever sent to the LLM).
• Exclude sources from profiling via exclude_sources.

Adding a new harness

Implement the three-method Adapter trait in src/adapters/ (name, discover, read), register it in registry() (src/sync.rs), add a fixture + test. discover returns fingerprints (mtime^size or row versions); the sync engine handles all incrementals, writing, and git.

Development

See FEATURES.md for the roadmap (near-term: integrations only) and AGENTS-MAINTENANCE.md for the maintenance guide.

cargo test          # full suite: adapters, importers, pipeline, e2e (mock LLM, no network)
cargo clippy -- -D warnings