Manifesto · 2026

The cult of productivity sold you more inboxes.

Bella sells you fewer. One. The one that knows your board agenda, your last conversation with Markus, and which calendar conflict actually needs you — before you ask.


01

Persistent memory.

Not a session. A relationship.

Every AI tool you have tried reset when the tab closed. It forgot your last meeting, your colleague's name, the commitment you made on Thursday. You became the memory layer — pasting context, re-explaining history, carrying the cognitive load yourself.

Bella remembers across every channel, year over year. Your calendar history, your email threads, your meeting transcripts, your relationship signals — all indexed, all retrievable, all feeding every action Bella takes on your behalf. When Bella drafts a follow-up to Sarah Chen, it knows the last three conversations, the outstanding commitment from Q3, and the tone she responds to.

Persistent memory is not a feature. It is the foundation. Without it, every AI interaction is your first.

0days of context, always on

02

Cascade routing.

96% of queries resolved without burning compute you don't need.

Most AI systems send every request to the most powerful — and most expensive — model available, regardless of whether the task requires it. 'What time is my 3pm?' does not need Opus. Rerouting a calendar conflict probably does not either. Drafting a board-level communication brief might.

Bella classifies every task before routing it. Simple retrievals and structured lookups resolve locally or against a crystal cache — sub-100ms, zero marginal cost, zero data exposure. Mid-complexity reasoning routes to Claude Haiku. Only genuinely complex synthesis — multi-document analysis, communication DNA reports, ambiguous autonomous actions — escalates to Sonnet or Opus.

The result is a system that feels fast because it is fast, that costs what the task is worth, and that keeps sensitive reasoning as close to your data as the task allows. Intelligence is not one-size-fits-all. Routing should not be either.

0%of queries resolved at Haiku tier or below

03

Scoped autonomy.

Five levels. A kill switch on every surface. You decide where Bella ends and you begin.

Autonomy without constraint is not intelligence — it is liability. Every agentic system that acts on your behalf is making a risk judgement. The question is whether you made it, or the system did.

Bella classifies every action across five tiers: read-only observation, summarisation and synthesis, draft-and-confirm, execute-with-notification, and fully autonomous execution. Each tier requires explicit activation. Each surface — email, calendar, files, groceries, vault — carries its own autonomy setting. Tier 4 on your shopping list does not imply Tier 4 on your outbox.

HMAC-signed confirmations gate every irreversible action. The kill switch is a single command, zero latency, across all surfaces simultaneously. You are not handing Bella the keys. You are giving it a precise, revocable, scoped mandate — and you can narrow that mandate at any moment, from any device, without ceremony.

0autonomy tiers, HMAC-confirmed

Every tool you added promised to simplify. Every inbox, every app, every AI assistant asked you to learn its language and manage its limitations. Bella inverts that contract. You set the level of autonomy. Bella does the rest.

Bella does it better.

Request access

No credit card required · Free trial on every plan