Seiche: funding and FX evidence for Python and AI agents

Research US-dollar funding, money markets, foreign exchange, and capital-market context with source dates, provenance, coverage, and known gaps attached.

Open the hosted terminal · Connect an agent · Read dated research answers

Seiche is an open-source money-, foreign-exchange, and capital-market evidence terminal, with its deepest live coverage in US-dollar funding. It joins public and official sources from the Federal Reserve, New York Fed, OFR, US Treasury, CFTC, and other authorities into a source-clocked research surface.

Outputs retain their as-of dates, provenance, evidence status, known gaps, and point-in-time eligibility. Use them for financial research, agent context, and reviewable evidence. Seiche does not execute trades or provide real-time quotes or investment advice.

• Live terminal: seiche.info
• Developer guide and tool runner: seiche.info/developers
• Source and full documentation: github.com/beepboop2025/seiche
• Hosted MCP server: https://api.seiche.info/mcp
• Cache-only Trade Safety context: https://api.seiche.info/api/trade-safety/risk-context

Country funding references are available through money_market_context with section="countries" (catalog) or section="countries", country="JP" (country observations). The REST equivalent is /api/v2/country-funding/JP. Coverage is partial: monthly references, daily curves, shared euro-area policy and national sovereign issuers retain their own identities and dates. See the country methodology.

Connect an AI agent without installing Seiche

Connect an MCP-capable client to the public Streamable HTTP endpoint:

https://api.seiche.info/mcp

For clients that accept mcpServers JSON configuration:

{
  "mcpServers": {
    "seiche": {
      "url": "https://api.seiche.info/mcp"
    }
  }
}

Public evidence tools can be used without an account, subject to service limits. Start with tools/list to discover the tools available to your caller, then use:

For example, this anonymous read returns a bounded funding summary:

curl --fail-with-body https://api.seiche.info/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  --data '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"money_market_context","arguments":{"section":"summary"}}}'

Keep returned dates, coverage, caveats, and unavailable or restricted states in your agent's answer. A successful request does not make every source current. See the MCP integration guide for Claude, Codex, Cursor, VS Code, and raw JSON-RPC examples.

Install the Python package for local research

Seiche requires Python 3.12 or newer. On macOS or Linux:

python3.12 -m venv .venv
.venv/bin/python -m pip install --upgrade seiche
.venv/bin/seiche --help

The package installs the seiche research CLI and seiche-mcp stdio server. To connect an agent to your local installation's public tool surface:

SEICHE_MCP_PUBLIC=1 .venv/bin/seiche-mcp

Local tools use local source and cache state. For an optional local collection and brief:

.venv/bin/seiche pull
.venv/bin/seiche brief

The first pull can be slow while public histories populate the local cache. Later reads reuse the cadence-aware cache; stale and unavailable inputs remain explicit. Local collection depends on upstream availability and source access.

Start the local REST API at http://127.0.0.1:8787 with:

.venv/bin/seiche serve

The PyPI wheel contains the Python backend, not the browser terminal. Use seiche.info for the hosted terminal. To serve a local browser UI, use a source checkout, install the backend there, and build the frontend with npm ci followed by npm run build in frontend/. The source checkout's API serves the resulting frontend/dist directory when it is present.

Connect the product family

Seiche provides funding and macro context. Follow a research question to the product responsible for the next evidence type:

Each product retains its own coverage, access rules, source clocks, and data rights. The shared connection guide lists the available hosted MCP endpoints.

Access and evidence boundaries

The public trade_safety_risk_context tool returns only the last completed Seiche regime, index, coverage, source-status counts, and conservative clocks. It is metadata-only derived context, never an executable or real-money-eligible claim, and does not inspect the attestation ledger. Stream attestations are verified separately and never confer per-order execution authority.

Sixteen evidence/context tools are anonymous: the current funding-stress read, cache-only Trade Safety context, historical analogs, public backtest, data health, crypto record, institutional flows, oil/funding, FX/materials, US money markets, world markets, connected research (research_network), structured FX and China (market_workbench), the latest article, India–UAE funding and gold (gift_city_context), and explicit gold financing scenarios (gold_inventory_carry). Five higher-cost forecast, replay, positioning, prose, and LLM tools require an account and are omitted from anonymous tools/list responses.

Authenticated hosted MCP adds five private Agent Room preview tools for identity-to-key binding, room creation, signed append, verified listing, and full-chain verification. The same protected operations have REST twins under /api/agent-room/. They are discussion and evidence primitives only: every record is non-executable, x402 cannot unlock them, and local stdio has no bearer principal with which to expose them. See the Agent Room guide.

The canonical registry ownership marker is:

mcp-name: io.github.beepboop2025/seiche

Data and licensing boundary

The source code is licensed under AGPL-3.0-or-later. Upstream data keeps its own terms: Seiche's software license does not grant redistribution rights over data from third parties. Public outputs distinguish redistributable observations, bounded derived context, metadata-only sources, restricted inputs, and declared gaps. Review the source citation and rights fields before republishing data.

For research citation, use CITATION.cff; for defects or data questions, open a GitHub issue.