dynamicfeed-tools 0.1.0

Live, verifiable data tools for AI agents — Dynamic Feed client + ready-made adapters for LangChain, CrewAI, LlamaIndex and Pydantic AI.

dynamicfeed-tools

Live, verifiable data tools for AI agents — a tiny typed client for the Dynamic Feed API plus ready-made adapters for LangChain, CrewAI, LlamaIndex and Pydantic AI.

Dynamic Feed serves 56 live-data tools across 18 verticals (weather, hazards, security/CVEs, space, sanctions, macro, shipping, ...). Every JSON response is Ed25519-signed, and every datapoint carries a provenance envelope — source, licence, timestamp and age — so your agent can cite where a number came from and you can verify it cryptographically.

pip install dynamicfeed-tools                      # client only (httpx is the sole dependency)
pip install 'dynamicfeed-tools[langchain]'         # + LangChain adapter
pip install 'dynamicfeed-tools[crewai]'            # + CrewAI adapter
pip install 'dynamicfeed-tools[llamaindex]'        # + LlamaIndex adapter
pip install 'dynamicfeed-tools[pydantic-ai]'       # + Pydantic AI adapter

Keyless first

You can start with no API key and no signup:

from dynamicfeed_tools import DynamicFeed

df = DynamicFeed()  # keyless

df.nearby_hazards(lat=-33.86, lon=151.21, radius_km=300)   # signed live-hazard scan
df.earthquakes(min_magnitude=5.0)                          # USGS, worldwide
df.reality_check("the latest Python is 3.12")              # anti-hallucination verdict
df.attest("wind_kmh", "gte", 120, lat=25.76, lon=-80.19)   # signed MET/NOT-MET attestation
df.receipt({"decision": "loan approved", "model": "gpt-4o"})  # signed compliance receipt

The rest of the curated surface (weather, forecasts, satellites, sanctions, CVE lookups) takes a free key — one POST, no card:

key = DynamicFeed.signup()           # or: curl -X POST https://dynamicfeed.ai/signup
df = DynamicFeed(api_key=key)        # or set the DYNAMICFEED_API_KEY env var

df.current_weather("Sydney")
df.weather_forecast("Tokyo", days=5)
df.satellites(group="stations")
df.sanctions_screen("ACME Trading Ltd")
df.check_vulnerability("lodash", version="4.17.10", ecosystem="npm")

If a key-gated method is called without a key, it raises MissingAPIKeyError with the signup instructions — it never sends a doomed request.

Want all 56 tools keylessly? Dynamic Feed also speaks MCP (https://dynamicfeed.ai/mcp, streamable HTTP, no key) — ideal when your framework has an MCP adapter. This package is the zero-infrastructure REST path with a curated, signed subset.

The curated tools

Tool	What it answers	Key?
nearby_hazards(lat, lon, radius_km)	live hazards around a point (quakes, fires, storms, disasters), signed	keyless
attest(metric, op, threshold, lat, lon)	signed parametric trigger: is wind_kmh >= 120 here, right now?	keyless
receipt(data)	Ed25519-signed, timestamped record of what an AI was told/asserted	keyless
earthquakes(min_magnitude, limit)	recent worldwide earthquakes (USGS)	keyless
reality_check(claim)	fact-check a claim against live data — never guesses	keyless
current_weather(city)	live weather anywhere	free key
weather_forecast(city, days)	up to 16-day forecast	free key
satellites(group, limit)	constellation catalog + physics latency floor	free key
sanctions_screen(name)	live US OFAC SDN + UK Sanctions List screen	free key
check_vulnerability(package, version, ecosystem)	full OSV.dev advisory lookup	free key

LangChain / LangGraph

# pip install 'dynamicfeed-tools[langchain]'
from dynamicfeed_tools.langchain import get_tools

tools = get_tools()                    # keyless tools only
tools = get_tools(api_key="dyn_...")   # all ten

from langgraph.prebuilt import create_react_agent
agent = create_react_agent("openai:gpt-4o", tools)
agent.invoke({"messages": "Any hazards within 200 km of Sydney right now?"})

CrewAI

# pip install 'dynamicfeed-tools[crewai]'
from crewai import Agent
from dynamicfeed_tools.crewai import get_tools

analyst = Agent(
    role="Risk analyst",
    goal="Ground every claim in live, signed data",
    backstory="Works only from verifiable sources.",
    tools=get_tools(api_key="dyn_..."),
)

LlamaIndex

# pip install 'dynamicfeed-tools[llamaindex]'
from llama_index.core.agent.workflow import FunctionAgent
from dynamicfeed_tools.llamaindex import get_tools

agent = FunctionAgent(tools=get_tools(), llm=llm)
await agent.run("Fact-check: 'the latest Node.js LTS is 20'.")

Pydantic AI

# pip install 'dynamicfeed-tools[pydantic-ai]'
from pydantic_ai import Agent
from dynamicfeed_tools.pydantic_ai import get_tools

agent = Agent("openai:gpt-4o", tools=get_tools())
result = agent.run_sync("What's shaking? Biggest earthquake in the last day.")

Which tools does an agent get?

All four get_tools() adapters behave the same way:

• no key → only the keyless tools (an agent is never handed a tool that is guaranteed to fail),
• key present (argument or DYNAMICFEED_API_KEY) → all ten,
• include_key_gated=True/False overrides the auto behavior.

Verifying responses

Responses are signed by default with the Dynamic Feed Ed25519 key. To verify:

1. fetch the public keys from https://dynamicfeed.ai/.well-known/keys,
2. remove the signature field, canonicalize the rest (JSON, sorted keys, compact separators), and check the Ed25519 signature over those bytes.

The full DF-VERIFY/1 spec lives at https://dynamicfeed.ai/standard, and the dynamicfeed-verify package does the verification in one line. Receipts and attestations can additionally be anchored to Bitcoin via POST /v1/anchor for independent proof of when.

Provenance envelopes

Datapoints carry their provenance — typically source, licence, measured_at / timestamp and age_seconds — so an agent can cite the upstream (USGS, NASA EONET, Open-Meteo, OFAC, OSV.dev, ...) rather than asserting numbers from nowhere. The authoritative source catalog (publisher, licence, commercial-use status, cadence) is keyless at GET https://dynamicfeed.ai/v1/sources.

Errors

DynamicFeedError
├── MissingAPIKeyError   # key-gated method, no key configured (raised client-side)
└── APIError             # non-2xx from the API (status_code, detail)
    ├── AuthError        # 401 / 403 — missing/invalid key server-side
    └── RateLimitError   # 429 — daily quota reached

Development

python3 -m venv .venv && . .venv/bin/activate
pip install -e . pytest
pytest

Tests run fully offline (httpx.MockTransport) and need none of the framework extras; adapter construction tests activate automatically when a framework is installed.

License

MIT — see LICENSE. Data licences vary per upstream source and are declared per datapoint and in the source catalog.