One platform. From free daemon to governed fleet.

The Crux Daemon is free, local, and genuinely useful on its own. The CueCrux Platform adds reach and rule — one deliberate step at a time. Here is every main feature, and exactly when each tier earns its keep.

Free forever · £0/user · mo · up to 10 agents

The Crux Daemon — the whole memory engine, free

This is not a trial. The local daemon is the acquisition, not the upsell: source-available, CPU-only, no account, no telemetry. Everything below runs on your machine and stays there.

Memory core

A versioned, append-only fact store with freshness decay, BM25, and bring-your-own dense retrieval.

  • Append-only BLAKE3 event store
  • Versioned facts + decay horizons
  • BM25 + BYO embeddings (uncapped)

Trust & receipts

Ed25519 CROWN receipts on every state mutation, verifiable offline — enforcement is a property of the wire, not a crippled binary.

  • Signed CROWN receipts
  • Offline verification (CLI)
  • RCX capability tokens (self-issued)

Continuity

Sessions that survive restarts and ExecPlans that carry work across runs, with token burn attributed per plan.

  • Checkpoints + handoffs
  • ExecPlans + live work board
  • Per-plan token attribution

Decisions & constraints

Record why a decision was made, declare boundaries in natural language, and check actions before they run.

  • Append-only decision records
  • Natural-language constraints
  • Pre-action pass / warn / block

Identity & custody

Agent passports, GDPR erasure, and self-certifying .cruxpack export — your data follows you.

  • 5-tier agent passports
  • Scoped forget (GDPR)
  • .cruxpack export + offline import

The console

A local operator console at localhost:14800 — the fleet, gate queue, cost, and your memory — no account.

  • Overwatch + Canvas
  • Documents + Explorer
  • Three themes (glass / dark / light)
Pro · £19/user · mo

Sync your memory across the fleet

You will know when you need this: the moment a second machine, or a second agent, needs the same memory. Pro takes what runs locally and makes it shared — without giving up the receipts.

The daemon binary does not change; a hosted capability token unlocks the shared backends. Downgrade any time and your local memory keeps working, untouched.

  • Sync across devices + agents (multi-device passport)
  • Hosted retrieval — better-dense rerank + extraction
  • Freshness alerts + dashboards
  • A monthly credit grant to start — 1,000 Cr
  • Everything in the free daemon, unchanged
Start Pro
  • Reconstruct any incident — evidence you, your customer, and an auditor can independently inspect
  • Fleet-wide cost attribution — who spent what, on which plan
  • A receipts timeline across every agent
  • Policy gates + human approvals on the actions that matter
  • Revocation that propagates across the fleet, with a signed record
  • A monthly credit grant — 5,000 Cr
  • Regulator-shareable audit packs (EU AI Act 2027 track / DORA)
Govern your fleet
Governance · £49/user · mo

Reconstruct any agent incident, on demand

When agents act on behalf of an organisation, someone has to be accountable when something goes wrong. Governance turns the fleet's receipts into an incident record — evidence you, your customer, and an auditor can independently inspect — plus the attribution, policy gates, and revocation to act on it.

It builds on the memory core; it is the second-tier control, not the starting point. If you do not need it, you do not pay for it.

Credits · ~£0.01/Cr

Meter only the heavy lifts

Some work needs a GPU or a frontier model. Credits pay for exactly those, per use — no subscription, no obligation. The local dense lane is never metered; credits buy the step-up, not your own machine.

Better-dense rerank + GPU fused retrieval

when local CPU cosine is not enough

LLM entity / relation extraction

fill the graph from raw content

Proofs, attestations, C2PA

publicly-trusted, cross-provider evidence

Hosted storage

beyond your own machine

Why memory beats brute-force context

At scale, stuffing everything into the context window makes strong models worse and costs more. Measured by ScoreCrux, our open agent benchmark.

Core recall at 2M-token scale

higher is better

Find 25 decisions + 5 needle facts across 3,346 documents — the Delta corpus, ResearchCrux.

28%
Stuffed model A · from 44%
8%
Stuffed model B · from 28%
80–100%
Tool-mediated memory retrieval

More context makes strong models worse — and costs up to 6.6× more per run for far worse recall.

Context tokens for the same answer

lower is better

Near-identical score; a fraction of the token bill.

Crux · 95/1001,263
Vendor-native memory · 97/1008,428

~6.7× fewer tokens, and most accurate at resolving which fact was current — ScoreCrux context benchmark, claude-sonnet-5 arm.

Compare the tiers

Everything free stays free at every tier. Credits are pay-as-you-go on any plan.

CapabilityFreeProGov.
Local memory, receipts & console
100+ MCP tools
.cruxpack export — custody, no lock-in
Sync across devices + agents
Hosted better-dense rerank + extractioncreditscreditscredits
Freshness alerts + dashboards
Monthly credit grant
Fleet-wide cost attribution
Policy gates + approvals
Revocation that propagates, with a signed record

Upgrading is low-risk by design

Source-available

Read every line (CCL-1.0); each release converts to Apache 2.0 three years after it ships. Verify your own enforcement.

Custody, always

Export everything as a self-certifying .cruxpack any daemon can re-ingest. Your memory survives us disappearing.

Upgrade when it's obvious

No lock-in and no obligation. Downgrade and the local daemon keeps working. Pay only when the value is undeniable.

Start where it costs nothing

Run the daemon locally, free forever. Upgrade the day a second machine needs the same memory.