AI you can trust
AI you can trust is Altretta's flagship difference: the model answers from your notes, quotes the exact source, and tells you honestly when the evidence is thin. Altretta ships no LLM of its own — you bring your own model and your own subscription, and Altretta turns your vault into a grounded, cited knowledge source it can read.
Obsidian and Logseq leave AI to you: paste your notes into a chat window and hope. You can't see what the model actually used, you burn tokens re-sending your whole vault, and nothing stops it from inventing a decision you never made. Altretta closes all three gaps — every answer is cited, only the relevant passages are sent, and each answer is held to an honest groundedness verdict.
Why this is different
A normal AI chat about your notes has three problems. Altretta solves each one.
| The problem with "paste and pray" | What Altretta does |
|---|---|
| You can't tell what the model read | Every claim cites its source as source:lines, e.g. [decisions/2026-03-db.md:12-19] |
| It fills gaps with confident guesses | A groundedness verdict (grounded / weak / ungrounded) calls out thin or missing evidence honestly |
| Re-sending the whole vault burns tokens | The AI queries first, sending only the handful of passages that bear on the question instead of the whole vault |
The query-first idea
Instead of stuffing your notes into a prompt, Altretta exposes your vault to any MCP client. Before the model answers, it queries the vault for the exact passages that bear on your question, and each passage arrives with a signed provenance anchor — proof of what the note said and when. The model is instructed to answer only from that evidence, cite it, and admit when it isn't enough.
Start here
Connect your AI
One click to detect your MCP client, mint a scoped access token, and wire it up — Claude Desktop, Claude Code, Cursor, VS Code and Windsurf.
Grounded, cited answers
Ask questions inside the app or from your AI client and get answers with source:lines citations and a groundedness verdict.
AI profiles
Bring your own model — Anthropic, OpenAI, OpenRouter, Gemini, DeepSeek, Groq, Mistral, xAI, Z.AI, local Ollama, or any OpenAI-compatible endpoint.
What your AI sees
A transparency panel showing the context exposed to the model — and the controls that decide what it may read and write.
Activity log & consistency review
A log of everything the AI queried, and a consistency review that flags notes contradicting the rest of your vault.
Review inbox: approve proposed notes
A staging area where a model can propose a new note for you to approve or reject, instead of writing it straight in.
New to Altretta's AI? Do them in order: Connect your AI first, then Grounded, cited answers.
Related
Verifiable knowledge
The signed history that makes every AI citation provable — the other half of "AI you can trust."
Provenance & proofs
The signed provenance anchor behind every grounded answer, inspectable in the Proof Inspector.
The semantic graph
The graph the query-first AI reads to pull only the passages that bear on your question.
Ask questions about your code
The same grounded, cited AI answering from an ingested codebase, not just notes.
MCP tools reference
Every tool the AI uses to query your vault, listed and explained.
Glossary
Grounding, provenance, DAG, MCP — the terms behind Altretta's AI, defined.