How it works
How Datailor learns your defaults
Datailor runs locally, reads what you have already told AI, and turns repeated corrections into preferences every agent can use.
Local by default
Datailor starts from your machine, not from someone else's cloud. Your preference library is a local, editable file that you can inspect, move, or delete.
It reads what you have already told AI and turns repeated corrections into defaults. Optional model backends are only used when you configure them.
pipx install git+https://github.com/fyaic/bondie-preference-MCP.git
Cold start from where you already work
When your preference library is empty, Datailor looks for the places where your AI conversations already live. It does not ask you to write rules from scratch.
The first scan extracts preferences you have already expressed: how you like code reviewed, when tests should run, what writing style you avoid, and when an agent should proceed quietly.
Current local history import covers Claude, Codex CLI, and Kimi CLI. Cursor and Trae history importers are planned as their local storage support matures.
Learn from the moments that matter
Preferences are not only captured when you explicitly say "remember this." Datailor watches for the moments where your intent becomes clear.
| Moment | What Datailor learns |
|---|---|
| When you say it | Direct preferences you express during conversation |
| When you repeat it | Corrections that keep coming back across turns or sessions |
| When your actions prove it | Signals from completed work, such as tests, formatting, or review steps |
| When the session ends | Late patterns that only become visible after the whole conversation is seen |
Merge, resolve, then decide
New signals are not blindly added as more rules. Datailor compares them with your existing preference library, merges duplicates, and freezes conflicts for review.
Before an agent answers, Datailor checks whether any preference fits this exact task.
- Is it relevant? A writing preference should not affect a refactor.
- Is it confirmed? Stronger signals apply before weak guesses.
- Is it still fresh? Old defaults lose weight unless you keep using them.
Put the right default where the agent can use it
Datailor does not wait for you to paste rules into every chat. It places the right preference where the agent can actually use it, before the repeated question comes back.
| Path | How it helps | When it applies |
|---|---|---|
| Static rules | Written to AGENTS.md / .cursorrules | Agent startup |
| Session warm-up | Session-specific cache generated | Session initialization |
| MCP real-time | get_preference_decision called |
Before every agent response |
| Local fallback | fallback_rules.md read directly | When MCP is unavailable |
Static rules and local fallback keep your defaults available even when a client is less proactive. Real-time MCP calls make the preference layer sharper when an agent asks before responding.
// MCP tool called before every agent response
get_preference_decision({
agent: "claude",
context: { task: "refactoring", language: "typescript" }
})
// Returns: the relevant preferences for this task
You can see and correct the loop
Datailor is not a black box. Every learned preference and every injection decision is something you can inspect, confirm, reject, or rewrite.
- Confirm Mark a preference as right so it applies with more confidence next time.
- Reject Stop a preference from applying when Datailor learned the wrong lesson.
- Correct Rewrite the preference in your own words instead of accepting an automatic summary.
- Trace See why a preference applied, when it was injected, and which conflict is waiting for review.
The loop keeps improving
Your preference library is not static. Feedback changes what applies, how strongly it applies, and when Datailor should stay quiet.
- Confirmed preferences get stronger The defaults you approve become easier to apply in similar tasks.
- Rejected preferences stop applying A wrong guess is removed from the active path instead of nagging you again.
- Repeated corrections lower confidence If you keep correcting the same behavior, Datailor treats the old preference as unstable.
- Clean usage adds stability When a preference applies without needing correction, it gains a small amount of trust.
- Conflicts stay visible When two preferences disagree, Datailor pauses automation and asks for a human decision.
Your defaults travel with you
Datailor is not tied to any single AI product. Your defaults stay readable, portable, and useful across machines, agents, and workflows.
Run it live
Run Datailor locally and let MCP-capable agents ask for the right preference at the right moment. Install once, then keep working as usual.
Take it with you
Export your preferences as a small Markdown file. Move them to another machine, another agent, or another workflow. They belong to you, not to any platform.
Install once. Use everywhere.
pipx install git+https://github.com/fyaic/bondie-preference-MCP.git