OpenChart
Free Desktop charting with Claude or Codex. Describe alerts, research chart moves, and choose your data providers.
AegisAgent gives autonomous AI agents restricted authority to spend USDC through an owner-controlled vault. Owners define spending limits, authorized recipients and expiration rules, and can revoke pe
Translation pending
FinanceGitHub @Lakshmikanth-3OnlineOwner unconfirmed
Have you tried it? One tap is enough. No sign-in needed.
Repository information has not been collected yet.
No updates detected yet. GitHub releases will appear here.
Enterprise treasury agent: reconciles each invoice across supplier, ERP and treasury data in Zetaris, shares evidence through the Meterless World Model, and lets NVIDIA Nemotron (via Nightshift) decide only inside a deterministic payment policy
Taken from the repository README
Loading the full README.
Read the full README ↗Live app: https://aegisagent-ai.vercel.app · Treasury review (sample mode on the hosted site; live Zetaris + Nightshift locally) · Owner console · Public repository
Autonomous Enterprise Treasury & Policy Guardrail Agent
Open Agent Hackathon 2026 · Wildcard [Tinkerer]
AegisAgent reconciles invoices against independent supplier, ERP and treasury exports through Zetaris, shares evidence through the Meterless World Model, and asks NVIDIA Nemotron through Nightshift to reason inside a deterministic payment ceiling. The policy critic can request another lookup before allowing a decision.
Requires Node.js 22 and pnpm 9.15.9. From a fresh clone:
cd app
pnpm install --frozen-lockfile
npm run treasury:sample -- ../input_examples/example_1.json
npm run treasury:sample -- ../input_examples/example_2.json
npm run treasury:sample -- ../input_examples/example_3.json
npm run treasury:sample -- ../input_examples/example_4.json
These produce PAY_EARLY, HOLD, REJECT and ESCALATE respectively. Outputs are recomputed from raw cached records; the workflow never reads output_examples/. Try changing the amount or PO number. Sample mode makes no sponsor requests and never submits a transaction.
For the UI, copy app/.env.example to app/.env.local (the template defaults SAMPLE_MODE=true), then run npm run dev -- -p 3001 from app/ and open Treasury review. Port 3000 belongs to the local Zetaris UI.
With the app running, from the repository root:
curl -X POST http://localhost:3001/api/treasury/run -H "Content-Type: application/json" --data-binary @input_examples/example_1.json
The API accepts invoiceNumber, supplierId, poNumber, amountUsdc (decimal string), discountBps, daysEarly, payoutAddress, and optional notes. Mode comes from server-side SAMPLE_MODE; clients cannot choose a live or cached data source. Results include the decision, structured audit, query evidence, real role handoffs, World Model audit, run ID and input hash. Validation errors return HTTP 400. Missing live evidence returns HTTP 503 with ESCALATE. No transaction is sent by this endpoint.
npm run treasury:bootstrap. It copies the supplied exports into the configured driver's shared data path and idempotently registers four file-store tables across AEGIS_REGISTRY, AEGIS_ERP and AEGIS_TREASURY. Default container: zetaris-driver-1. Override ZETARIS_DRIVER_CONTAINER and ZETARIS_DATA_PATH for your deployment. -- --skip-copy uses data already provisioned at that path.npm run treasury:run -- ../input_examples/example_1.json or use the same API/UI. Inspect actual returned model IDs and query records. Nightshift model discovery is available at /v1/models with the Runtime bearer token; meterless/auto alone does not prove a Nemotron request.The data sources are small synthetic enterprise exports, not real customer systems. Live mode executes real Zetaris queries over those exports; sample mode reads the same raw data locally. This distinction appears in the UI, audit and logs. Zetaris is required in the live assessment path: an outage escalates rather than silently substituting cached evidence.
PlannerAgent assigns reconciliation. ReconcilerAgent retrieves registry, PO, invoice-history and treasury evidence, then writes typed entities and facts to the World Model. PolicyCriticAgent reads that shared state, checks consistency, and requests a broader PO lookup when scoped evidence is missing. TreasuryReasonerAgent calls Nemotron through Nightshift with precomputed math and an allowed decision set. The critic independently validates the final decision against that ceiling. Retrieval and critique roles use deterministic code; the reasoning role uses the LLM in live mode. They are distinct workflow roles, not four independent model calls.
The actual Apache-2.0 World Model reference is vendored with attribution in app/src/lib/treasury/world-model/. Each assessment uses a private persistent namespace; the reconciliation, critique and reasoning roles write/read evidence and risk facts through it. Its event log and provenance are returned in the review. All eight official conformance checks pass via npm run world-model:test. The reference uses a single writer and local JSON storage, suitable for a local prototype rather than a distributed production store.
Every run creates app/logs/run-.jsonl and writes events to stdout. Traces include the input hash, roles, handoffs, query SQL/results, Engine activity, model request/response and policy validation. State lives under app/.data/.world/. AEGIS_LOG_DIR and AEGIS_STATE_DIR select other writable directories. Runtime traces/state are git-ignored. Sanitized, generated synthetic-data examples live in output_examples/ and logs/ for reviewers; they are evidence, never runtime inputs.
Treasury review is an assessment endpoint: it returns a decision with evidence and never verifies chain signatures or moves funds. Missing metadata, unavailable sources or conflicting evidence always prevent an autonomous approval.
From app/:
npm run treasury:test
npm run world-model:test
npm run earlypay:test
npm run typecheck
npm run build
The workflow and adapter tests cover all four decisions, changed inputs, cross-supplier PO discovery, source outages, input validation and hostile-model overrides. The existing EVM suites exercise real contracts locally, including randomized attacks. See the readiness report for recorded results and remaining submission actions.
From the repository root:
docker build -t aegis-agent .
docker run --rm -p 3001:3001 -e SAMPLE_MODE=true aegis-agent
For live mode use --env-file app/.env.local and set METERLESS_BASE_URL / ZETARIS_REST_URL to host.docker.internal instead of localhost on Docker Desktop. Docker does not install anything when the app starts. Provider tokens and local env/state are excluded from the image context.
The on-chain settlement layer (vault contracts, signatures, SDK, supplier/buyer flow and early-payment maths) was built before the build window and is declared as pre-existing work in the submission form. New work includes live Zetaris reconciliation, Nightshift provider, World Model integration, role feedback, structured assessment API/UI, cached-data judging mode, generated traces/examples, Docker packaging and submission materials. Submission text and demo script describe the new work.
MIT