ForgeRelay

ForgeRelay is an AI-assisted RFQ clarification service for small manufacturers. It converts an incomplete request for quotation into a quote-ready, evidence-backed work package and coordinates the missing information with suppliers.

Demo video: https://youtu.be/cg_hcuHM5XQ

The project was created on July 26, 2026. It is a new, independent product and codebase built during the active submission windows for the competitions listed below. It uses synthetic sample data and does not contain customer documents or code from pre-existing LinkSea or Dollar products.

The three-minute loop

1. Paste an incomplete manufacturing RFQ.
2. Gemini extracts requirements, uncertainties, and missing commercial facts.
3. DataHub MCP adds ownership and lineage context before the agent recommends action.
4. ForgeRelay produces a prioritized clarification plan.
5. After an exact, content-bound safety preview and explicit human approval, CALL-E contacts the supplier.
6. ForgeRelay re-fetches the authoritative CALL-E record and converts grounded transcript spans into answered, unknown, or conflict outcomes.
7. CockroachDB stores the minimum structured memory required for another agent to resume the case on AWS.

Event targets

Event	ForgeRelay proof
Build with DataHub	Metadata-aware RFQ change impact using DataHub MCP
Build with Gemini XPRIZE	New AI-operated small-business service on Google Cloud
CockroachDB × AWS Agentic Memory	Durable, resumable RFQ case memory
CALL-E	Real supplier clarification calls with structured outcomes

The CALL-E entry also has a validated public Agent Skill contribution in CALLE-AI/awesome-phone-call-agents#35.

Local start

npm install
copy .env.example .env
npm start

Open http://localhost:3000. DataHub GMS remains on its official local port, http://localhost:8080. Without credentials the app runs in a visibly labeled deterministic demo mode. A competition release is not considered complete until the relevant sponsor integration is configured and captured in the evidence pack.

Local DataHub proof

Start DataHub Core with the official quickstart, seed the synthetic catalog, and connect the official MCP Server over stdio:

# Run these in an isolated Python 3.11 environment.
python -m pip install -r scripts/requirements-datahub.txt
datahub docker quickstart --version stable
python scripts/seed-datahub.py

Set these local-only values in .env:

DATAHUB_MCP_COMMAND=uvx
DATAHUB_MCP_ARGS=mcp-server-datahub@latest
DATAHUB_GMS_URL=http://localhost:8080
DATAHUB_ENABLE_MUTATIONS=false

Then run npm start and npm run smoke. A successful live proof reports datahubSource: "datahub-mcp" and the tool trace search → get_entities → get_lineage. See submissions/datahub/EVIDENCE.md.

Live demo

The public, credential-free judging build is available at https://forgerelay.vercel.app. It uses the same synthetic fallback as the local app and labels sponsor integrations honestly when they are not configured.

Safety

• Live calls are disabled by default.
• The server rejects arbitrary CALL-E tasks. It builds the questions and approved disclosure context from the saved RFQ case, not browser-supplied free text.
• A masked preview shows the recipient basis, caller identity, exact questions, approved facts, and forbidden actions. Approval is bound to that content by a SHA-256 review hash.
• CALL-E idempotency is derived from the complete safety contract, so a retry cannot silently place a second call or reuse approval after an edit.
• Current CALL-E webhooks are treated only as untrusted wake-up signals. The server re-fetches GET /v1/calls/{id} with its API key before recording a result, and stores no full transcript.
• No source document is sent to a provider; only the minimum RFQ facts entered by the user are used.
• CockroachDB stores structured facts and evidence references, not uploaded documents or secrets.
• DataHub mutations are off by default.

License

Apache-2.0. See LICENSE.