Project Kintsugi.
fungus reparator — values as architecture, not prompts
An ethical agent framework where organizational values are hard constraints, not suggestions. Memory that persists across agents and years. Security with independent veto authority. The gold in the cracks is structural.
Your values are not suggestions.
Most AI agent frameworks treat ethics as a prompt. Write "be helpful and harmless" in the system message and hope for the best. When the agent encounters a conflict between helpfulness and a value, the prompt loses. Kintsugi doesn't prompt values — it enforces them. Hard constraints, not soft suggestions. If your organization says "never share member data with third parties," the Shield layer blocks that action. No override. No appeal. No creative reinterpretation.
Security with teeth.
Six independent security components. No soft overrides.
PII Redaction. Social security numbers, credit card numbers (Luhn-validated), email addresses, phone numbers — detected and redacted before storage. Not flagged. Redacted.
Intent Capsule Verification. Every agent action is wrapped in an intent capsule — a structured declaration of what the agent intends to do and why. The capsule is verified against the organization's values before execution.
The Shield. Four independent enforcers — budget, egress, rate-limit, circuit-breaker — compose a single verdict. If any enforcer blocks, the action is blocked. One veto kills the action. No consensus. No weighted average.
Invariant Enforcement. Mathematical constraints on system state that cannot be violated. If an operation would break an invariant, the operation fails.
Values that persist across every transition.
The BDI (Belief-Desire-Intention) framework encodes an organization's values as a machine-readable document. Beliefs — confidence-scored, source-tracked, verification-dated. Desires — measurable outcomes, not aspirational language. Intentions — specific, time-bound, resource-allocated. Principles — hard ethical constraints that the agent cannot violate regardless of what a user requests.
Four organizational templates: Mutual Aid · Nonprofit 501(c)(3) · Cooperative · Advocacy. VALUES.json is hot-reloadable. Change your values, the agent changes its behavior. No retraining. No redeployment.
Architecture basis: BDI framework (Bratman, 1987; Rao & Georgeff, 1995) adapted for organizational agents. Values-as-constraints extends constitutional AI (Bai et al., 2022) to the action level.
Skills for communities.
Core Operations: Finance assistant · grant hunter · impact auditor · institutional memory · content drafter · volunteer coordinator
Community Aid: Mutual aid coordinator · crisis response · rapid response · resource redistribution · housing navigator · know your rights · food access · solidarity economy · coalition builder · community asset mapper
Each skill operates under the full security and values stack. A grant hunter cannot recommend funding sources that violate principles. A housing navigator cannot share client data beyond what the values document permits.
Start where you are.
| Seed | SQLite · no cache | Laptops, small orgs |
| Soil | PostgreSQL · no cache | Growing orgs |
| Grove | PostgreSQL + pgvector · Redis | Production, multi-agent |
Every tier includes the full security and values stack. The difference is performance and scale, not protection.
What makes this different.
Values are architecture, not prompts. The BDI framework, the Shield, the intent capsules — these are code that executes, not text that the model reads and might ignore.
Security is independent, not layered. Six components, each with independent veto authority. They don't need to agree. Any one can stop an action.
Memory is institutional, not conversational. CMA doesn't store chat logs. It extracts facts, scores significance, and maintains organizational knowledge that persists across agents, across sessions, across years.
Open source. Values-gated. The code is open. Commercial support is available only to organizations whose values align with ours.
Built by Liberation Labs — a worker-owned cooperative. AI team members alongside human engineers.