Back to projects

PROJECT / SYSTEMS

cupel

An AI agent team, run like production.

cupel runs our agents on our own machines with hash-locked approvals, an encrypted vault, guaranteed message delivery, hard spend caps and crash-safe restarts. One control panel shows all of it live.

cupel in 40 seconds. Sound on, captions in the player.

01 / 06

One team, many specialists.

An orchestrator routes every task to the specialist built for it. Each agent runs with its own role, its own permission profile and its own chat bot, on your own hardware.

Out of context or crashed, an agent is restarted by cupel and continues from its saved state. Idle agents pull the next planned card on their own.

02 / 06

Every handoff lands.

Agents talk over a message bus with a delivery guarantee. A message stays open until the receiver closes it, is redelivered when nobody acts, and is dead-lettered and flagged if it still fails.

In typical agent frameworks a handoff lives inside one run by default; durable state is an extra you wire up yourself.

  1. Orchestrator → Coder

    task: fix bug

    deliveredclosed

  2. Researcher → Orchestrator

    found 3 sources

    deliveredanswered

  3. Monitor → Orchestrator

    alert: disk almost full

    redeliveredclosed

03 / 06

Work you can see.

Every task lives on one live board, from planned to done. The maker can't sign off its own card: another agent or the owner accepts it, and the board enforces that rule.

In typical agent frameworks review is an optional add-on, and no rule stops a maker from approving its own output.

Made by the Designer. Its own sign-off is refused.

04 / 06

Runs while you sleep.

Routines fire on time in your local time zone. cupel's own daemon runs them whether anyone is watching or not, and the next runs are always on the control panel.

00061218 24 h local time
  • Memory cleanup
  • Morning briefing
  • Inbox triage
  • Health check
  • Daily summary
  • Backup & audit
  • Evening report

An example day, fired by the daemon.

05 / 06

You stay in control.

Every risky action stops at the gate. You approve exactly that command, the approval is locked to the exact command, and the very same command runs: nothing can be swapped in between.

Keys sit in an encrypted vault. Agents call outside services through a gateway that uses the key for them, so the key never lands in a prompt or a log. Paid API calls have daily and per-project money caps, and every external call is logged.

  • Encrypted vault
  • Money caps on paid API calls
  • Every external call logged

06 / 06

Every lesson, found again.

Skills, memories and cards are embedded into a vector index by a model on your own machine, so nothing leaves it. Search ranks them hybrid: vector similarity together with keyword match. Measured: the right skill in the top five for 16 of 17 test questions.

Every message from a person is looked up in that index. The best matches reach the agent with the message within 1.5 seconds, without the agent asking and without ever holding it up.

One memory serves the whole team, in hot, warm, cold and shared tiers, plus an append-only daily log. Its full-text search ignores accents and case. Shared skills and each agent's own are listed on the control panel, every skill load is logged per agent, and a health check measures dead links, orphan notes and the same lesson filed under two names.

cupel next to a typical agent framework

API keys
Typical: Environment variables
cupel: Encrypted vault, used through a gateway
Approvals
Typical: Opt-in per tool, off by default
cupel: On for every risky action, hash-locked
Spend
Typical: Turn, step or token limits
cupel: Money caps on paid API calls, per day and per project
Handoffs
Typical: Inside one run by default
cupel: Guaranteed delivery, redelivery, dead letter
Memory
Typical: Opt-in, vector search where offered, embedder set per project
cupel: On for the whole team, local embeddings, vector + keyword rank
Skills
Typical: SKILL.md files in some, use not logged
cupel: Shared and per-agent skills, every load logged
Review
Typical: Optional, no rule against self-approval
cupel: The maker can't sign off its own work

Typical: the documented defaults of four popular open-source agent frameworks, checked October 2026.

THE CONTROL PANEL

All of it on one screen.

Agents, board, schedules and machine health in one live panel: CPU, memory, swap and GPU at a glance. The status line on top changes colour the moment something needs a person.

The cupel control panel with demo data: an all is well line, the agents and their activity, the board totals, the machine monitor with CPU, memory, swap and GPU, and the next scheduled runs.
Shown with demo data.

YOUR AI TEAM

Coordinated. Reviewed. In control.

We build agent systems for businesses on the same foundations: clear roles, guaranteed handoffs, hash-locked approvals and hard spend caps.

TALK TO US ➔ AI SYSTEMS