Cognivern
The economic control plane for agentic work.
Cognivern makes autonomous work fundable: a business can define a bounded mandate, allocate capital to an agent, enforce what the capital may be used for, record what happened, and build the evidence needed to allocate more. Today Cognivern provides the governed execution and audit substrate; outcome accounting and external capital allocation are strategic next layers.
Try it: Live app · API · PromptOS
Decision models via Runware /v1/systemone: typesafe:jev@latest + runware:laya@1 (FREE until Oct 12) for confidence-gated spend.
Choose your path
Repository at a glance
src/backend/ Backend API, governance services, persistence, integrations
src/frontend/ Next.js dashboard and public product UI
packages/shared/ Types shared by the backend and frontend
contracts/ Solidity and Fhenix confidential-policy contracts
daml/ Canton/Daml sealed-bid model and sandbox scripts
tests/ Canonical backend, integration, and browser tests
tooling/scripts/ Development, demo, verification, deployment, and research tooling
examples/copilot/ Standalone Cognivern Copilot agent example
docs/ Architecture, operations, integrations, and project reference
Most contributors only need one or two of these areas. The repository contains production code, protocol experiments, deployment tooling, demos, and historical research in one place; you do not need to understand all of it before making a focused change.
Quick start
Docker (self-host)
One-command local stack (API + UI + Jaeger + seeded demo mandate):
cd deploy/self-host
cp .env.example .env
docker compose --env-file .env up --build
Open http://localhost:8080 — see deploy/self-host/README.md.
Local (pnpm)
Requirements: Node.js version from .nvmrc and pnpm 9.15.0.
pnpm install
cp .env.example .env
pnpm dev
In a second terminal, run the frontend:
pnpm frontend
For the public/demo path, most optional integrations can remain disabled. See the Contributing Guide for focused commands and the Developer Guide for configuration details.
Documentation
What it does today
Governed spend: agents propose, policy evaluates, wallets execute, evidence records — the agent never holds a key. On top of that substrate: confidential policy evaluation (Fhenix), confidential vendor selection (Canton sealed-bid rounds), verified intelligence rails (Telegraph/x402), and proof anchors (0G, X Layer). Integrations include Dynamic MPC wallets, KeeperHub, Cleanverse, and Flare — see the docs index for each rail.
Roadmap direction: funded mandates → attributable execution → outcome accounting → external capital allocation. Govern → attribute → measure → allocate.
Verification
Every pull request is verified against the live deployment. The suite is generated in the TestSprite portal against the running service, exported into this repo (21 modules across 13 API domains), and run on every PR by .github/workflows/testsprite.yml. Eleven failed run bundles are recorded locally under .testsprite/runs/ — gitignored, because run artifacts are bulky and machine-local; the portal is the system of record for runs, and CI re-uploads failure bundles as a 7-day artifact. The durable evidence in the repo is the exported suite and docs/history/LOOP.md, which cites a test id per iteration.
The suite is written to check that the service is telling the truth about itself, not just that the happy path returns 200. The service declares which endpoints are public (PUBLIC_API_PATHS) and the middleware enforces it — that declaration is a claim about the running service, and it went stale repeatedly. Seven of the first eleven verification iterations turned up that single class of bug: /api/governance/policies returning 401 while listed as public, /auth/* blocked while listed as public, /api/spendos/status 404 with no route handler, and /metrics/ux-events, /fhenix/encrypt, /mcp/governance-check and /projects/:id/* missing from the list entirely. The fix that ended the class was structural — isPublicApiPath() gained parameterised path matching instead of another one-off entry.
A happy-path smoke test finds none of those. What the suite covers, and what it does not, is in docs/TESTING.md; the iteration log is docs/history/LOOP.md.
License
MIT
