Agent infrastructure 1.1.0 · Released 15 October 2026

Give your trading agent evidence it can check.

Add funding, institution and liquidity research to your agent. Reuse MCP connections, diagnose setup and follow source-status updates with Python or JavaScript.

Install and check your connection Try a task without installing

Free anonymous public reads. No account or data API key required.

These tools supply research evidence and source metadata. The context stream contains no live prices. Public research endpoints cannot place orders; source publishers retain their data rights.

Start with one useful question

Build a repeatable research step.

Check funding conditions

Ask for a bounded USD money-market sample with source URLs, units and observation dates. Compare the same question on another research day.

Try a funding watch

Review a counterparty

Use LiquiLens for covered institutions, dated disclosures and evidence gaps. Preserve unavailable inputs when drafting a risk note.

Use an institution watchlist

Research exit liquidity

Use Undertow's indicative liquidity evidence alongside funding context. Keep venue coverage, missing fees and source restrictions attached to the result.

Review funding and exits

Python 3.10+ · Dedicated environment recommended

Install the released source, then run the doctor.

python3 -m pip install --force-reinstall 'financial-evidence[workspace] @ git+https://github.com/beepboop2025/financial-evidence-skills.git@agent-infrastructure-v1.1.0'
financial-evidence-trading doctor
financial-evidence-trading context
financial-evidence-trading context --events --limit 1

The source package still reports version 0.1.8. Keep the agent-infrastructure-v1.1.0 source pin and use --force-reinstall to replace an older installation.

The doctor makes bounded anonymous checks of four public research connections and the source monitor. It labels its requests synthetic. Its result separates connection success from freshness, restrictions and unknown coverage; a passing connection check does not qualify an input for trading.

Connection diagnostic guide · Full release and client guide · Pinned release source

Keep connections and evidence under your control

Use a client that fits your agent.

Python: reuse a product MCP session

One client context owns the session for a bounded job. This diagnostic example marks its calls as test traffic.

import asyncio
from financial_evidence.trading import ProductClient

async def main():
    async with ProductClient(mode="test", timeout=15) as client:
        first = await client.call("seiche", "data_health")
        again = await client.call("seiche", "data_health")
        print(first["timing"], again["timing"])

asyncio.run(main())

Repeated calls reuse initialization. Failed calls are not automatically replayed. Timing describes the client request, while each observation keeps its own source clock.

Session lifecycle and limits

JavaScript: follow source status

Download context-client.mjs beside your script. Use Node 22+ or a modern fetch runtime.

import {
  contextSnapshot, contextEvents
} from './context-client.mjs';

console.log(await contextSnapshot());
for await (const event of contextEvents({timeoutMs: 30000})) {
  if (event.event === 'reset') {
    console.log(event.data);
    break;
  }
}

The stream begins with the latest full state. Reconnects do not replay every update. Python's ContextClient provides the same snapshot and event contract.

Python context example and stream limits

Already have an MCP client?

Add this Streamable HTTP endpoint, discover its live tool schemas and begin with a read-only research task:

https://agents.seiche.info/products/financial-evidence/mcp

Client setup and downloadable configurations · LangGraph, CrewAI, OpenAI Agents and Pydantic AI recipes

Readable by people and agents

Check the source, even when the connection works.

Monitor age and source age differ. Recheck valid_until while holding a snapshot. A current monitor can report stale, restricted or unknown sources.
A heartbeat proves transport liveness. It does not refresh an observation or turn a daily reference rate into a tick quote.
Access and permitted use differ. Retain source URLs, units, observation dates, restrictions and missing values. Public access does not grant redistribution or execution rights.

Snapshots are prepared from the scheduled source monitor. Reading the snapshot or stream does not trigger an upstream market-data fetch.

Context contract · Current source-status snapshot · SSE endpoint: https://agents.seiche.info/v1/context/events

Read the complete delivery contract · Get help with a redacted diagnostic