Opportunity Radar 📡

Ask out loud what public funding you can actually apply for — and get an answer with real numbers, or none at all.

Every year billions in public grant funding goes unclaimed, not because people are unqualified but because nobody has time to read grants.gov on a Tuesday night. Opportunity Radar is a self-hosted MCP server that turns that database into something you can ask a question to: "any grants open for a small AI company?", "what closes in the next two weeks?", "am I even allowed to apply for that one?"

Because Alexa+ speaks MCP, the same server answers in the kitchen.

Built for the Build, Ship, Shape: Amazon Developer Hackathon — Alexa+ track, with the AWS Builder and Open Source mini challenges.

What makes it different: it refuses to guess

A funding answer that is confidently wrong costs someone a weekend. So the honesty rules are structural, not prompt-deep:

The test suite asserts each of those (tests/test_radar.py::HonestyUnderFailure).

Related MCP server: grants-gov

Tools

Run it

git clone https://github.com/kevin9327/opportunity-radar && cd opportunity-radar
python -m radar.server              # http://0.0.0.0:8080/mcp

No API key is needed: grants.gov search is public. Responses cache for 6h.

# point any MCP client at it
curl -s localhost:8080/mcp -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"deadline_watch","arguments":{"interest":"artificial intelligence","within_days":30}}}'

Fit ranking runs on your own machine

rank_for_me asks a model whether each opening is worth your Saturday. It prefers a local model over Ollama — no key, no account, and the sentence describing your situation never leaves the house. That matters when the query is "grants for a family caring for a disabled child".

ollama pull llama3.1:8b      # engine: local

Fallbacks are automatic and always labelled in the response: Amazon Bedrock if AWS credentials happen to exist, then plain keyword overlap so the radar still answers offline. Set RADAR_RANK_ENGINE=local|bedrock|overlap to pin one.

How it fits together

flowchart LR
    A["Alexa+ / any MCP client"] -->|Streamable HTTP · spec 2025-11-25| B["radar.server<br/>JSON-RPC + SSE"]
    B --> C["6 tools<br/>honesty gates"]
    C --> D["grants.gov<br/>search2 + fetchOpportunity"]
    C -->|fit score + reason only| E["Amazon Bedrock<br/>Claude Haiku"]
    C --> F[(6h disk cache)]

radar/server.py implements the transport by hand against the MCP spec — one endpoint, initialize handing out a session id, tools/list schemas generated from Python type hints, tools/call returning both spoken text and structuredContent. No framework in the way, which made the spec easy to read off the wire while debugging.

Tests

python -m unittest discover -s tests -v     # 13 offline
RADAR_LIVE=1 python -m unittest discover -s tests   # + 2 against grants.gov

CI runs the offline suite on every push and the live-source suite weekly, so a silent API change shows up as a red build rather than a wrong answer.

MIT © 2026 kevin9327