Seiche MCP Server

SEICHE

A seiche is a standing wave in an enclosed body of water — invisible from the shore, until it sloshes over the edge. Funding stress behaves the same way.

Seiche is a free, open source (AGPL-3.0-or-later) money-, foreign-exchange and capital-market evidence terminal organized around the dollar funding system — US money markets, 22 public H.10 currency reference series, the Treasury capital-market complex, the global basins connected through swap lines, and the offshore-dollar crypto basin moored to the T-bill market through stablecoins. The core US board and USD desk use free official/public APIs (FRED, NY Fed Markets, OFR STFM, Treasury FiscalData, CFTC, ECB Data Portal, DeFiLlama, Coinbase Exchange). The global market-pack catalog can also declare licensed or tenant sources, but public outputs obey each source's redistribution policy and never expose restricted raw values.

Seiche publishes a construction-PIT historical diagnostic—with misses and eligibility limits kept visible—and is now accruing a separate as-published forward record that can be verified instead of reconstructed. Licensed terminals provide much broader real-time data and execution workflows, while OFR and NY Fed dashboards provide authoritative source views. Seiche is a narrower, opinionated research layer: forward-looking, alerting-ready, and provenance-honest. It is not a Bloomberg/Reuters replacement, a real-time quote service, or an execution venue.

The lab

Seiche is one of three altitudes in a single liquidity lab that fills that gap: Seiche reads the plumbing, LiquiLens ranks the institutions standing on it, and Undertow prices market liquidity itself (who provides depth in each segment, what an exit at position size costs today). The wiring between the three is real, not a brochure: the desk assistant pulls the LiquiLens failure board over MCP, the MARKET tab renders Undertow's published pack live, and the Windfetch engine reads the FETCH pack built in the Undertow repo back into this board. On names: the internal composite engine called "undertow" (critical slowing down, engines/undertow.py) is unrelated to the Undertow sister product.

Use Seiche everywhere

Seiche is distributed as software, an MCP service, a research dataset, and machine-readable catalog metadata. Status is receipt-based: listed means a durable public record and receipt exist, prepared means an owner-controlled submission or publication remains, and blocked means a named prerequisite must clear first. “Usable in repo” describes local artifact readiness, not a ledger status. The auditable source of truth is distribution/submissions.csv.

Eleven evidence tools remain anonymous and free. Five compute-heavy tools are account-gated; client and catalog copy must preserve that boundary. See the distribution and container trust guide for verification and release invariants.

v2 "Deep Water" — an accountable engine fleet, layered analytics, one terminal

v2.3 "Letters of Marque" (built in tandem across two sessions) adds the forecast layer and the layer that makes every other layer accountable: Undertow (critical slowing down — the basin's damping, measured on ordinary days), the Swell Forecast (the funding-stress forward curve — P(pop ≥ x bp) by date, six weeks out, from the public forcing calendar), The Stack (walk-forward ensemble of every event forecaster — rule, ML, analogs, Swell — plus The Tell, with a disagreement gauge), The Book (HELM tab — explicit daily positions on 2y/10y duration proxies, S&P and BTC over a T-bill base, walk-forward P&L with costs, block-bootstrap Sharpe CIs and mandatory benchmarks, verdict printed even when it loses), a hash-chained as-published track record shipped inside the static publish (nobody, including the operator, can quietly rewrite a bad month), and the Far Basin — Palimpsest's censorship-fear channel (palimpsest.info), a policy confession signal no market data vendor carries, honestly quarantined until it accrues testable history.

★ = methods invented for this tool.

The desk assistant: seiche ask "why is the index elevated?" (or the Ask box on BOARD) answers strictly from a deterministic context pack of the live board — every number cited to its engine and as-of date, "not on the board" instead of improvisation. Routed through free-llm-router's free tiers, or any OpenAI-compatible endpoint via SEICHE_LLM_BASE_URL; with neither configured it returns the context pack itself.

The analytics layers on top:

• The Tell — plumbing percentile minus market-priced-stress percentile (VIX, HY/IG OAS, rates vol). Positive = the basin is sloshing and the screens haven't noticed. The whole thesis in one tradeable number.
• The Estuary — upstream FX/material pressure minus the funding pressure already priced in SOFR and commercial paper. A positive Passage gap says the cash burden is running ahead of its usual landing zones; the discovery/holdout map then shows which links actually survived unseen history. Monthly physical breadth is never filled forward into the daily evidence, and the cash lab never pretends user assumptions are observed exposures.
• The Navigator — an LLM forecaster made accountable: one committed P(funding event, 5bd) per data-day, grounded strictly in the live board, written into the hash-chained record. An LLM cannot be honestly backtested (it has read the history), so its FORWARD record is its only evidence and its weight stays zero until that record earns a hearing. seiche navigator.
• The Communiqué — FOMC statements read as vintage-stamped data: frozen deterministic lexicons score policy direction, balance-sheet bias and funding-stress vocabulary per statement; the change vs the previous statement is the signal, and the Time Machine replays text as it stood.
• The TED bridge — the ML Lab pretrains on the TED spread's 1990–2018 funding-stress record (2008/2011/2016) in the same feature slots, down-weighted, and publishes the transfer gain vs the SOFR-only model either way.
• The Stack + The Book — the rule index, ML Lab, Tide Tables analogs, the Swell curve and Bathymetry's first-passage odds all emit P(funding event, 5bd); the Stack calibrates and blends them walk-forward (publishing the equal-weight mean whenever the fitted blend can't beat it), publishes member dispersion as a first-class ambiguity signal, and the Book converts the result into explicit daily paper positions with costs, benchmarks and bootstrap CIs. Every view's daily forecast and every position is appended to the hash-chained PIT record — a track record no reconstruction can polish.
• Turn Barometer — forecasts the next month/quarter-end turn's severity with leave-one-out cross-validation, always benchmarked against a naive forecast. When the model can't beat naive, it says so and publishes naive instead.
• Playbook — what S&P/VIX/OAS/yields did the last N times the board looked like this, in native units, with n printed. Decision support, not advice.
• PROOF — the construction-PIT diagnostic lab: the index rebuilt with expanding-window statistics only (causal truncation is unit-tested, but source history is final/current-vintage), recall/precision with Wilson 95% intervals, run-level precision (alert days are serially correlated; runs are the honest trials), episode-by-episode lead times including the ones it missed — and the orthogonal signal test: the same event-capture with the target's own variable family (spread/tails) removed from the signal. At build: orthogonal recall 69% [CI 42–87] vs 62% full, with the structural components alone at the 98th–100th percentile 42 days before the Sep/Dec-2025 squeezes. The claim is causal structure, not autocorrelation.
• Time Machine — final/current-vintage construction-PIT reconstruction using observations dated on or before any date since ~2018; it does not recreate the publication vintage visible then. Reconstructed to Sep 12 2019, the board reads EROSION with reserves $576B below the kink and flags Sep 16 2019 — the exact day the repo market broke — as a crunch window.

Principles: no naked numbers (every value carries source + as-of + staleness), fail-loud (a dead feed shows as DEAD and reduces published coverage — it never silently vanishes), honest lags (COT is T+3 by construction; shown, not hidden), and honest scope (a registered market with no qualifying public benchmark stays visible as unavailable; a policy rate or restricted input is never relabelled as a live traded benchmark).

Public world-market contracts

• GET /api/v2/world-markets returns the versioned seiche.world-markets.v1 projection of an already completed snapshot. It connects money-market funding, 22 official daily FX reference series, three trade-weighted dollar indexes, and a bounded macro-capital transmission view through selected Treasury, credit, volatility, dealer, positioning and commodity evidence, plus a signed metadata-only China macro side projection over four release-reviewed NBS series identities. The response carries canonical citation URLs, separate generation/evidence clocks, and observed, derived, structural, restricted or unavailable status. Add ?section=forex (or summary, money_markets, capital_markets, china_macro, sources, methodology, all) for a smaller projection. The request never starts collection, scans historical repositories or fits a model.
• Crawlable citation pages live at /markets/, /markets/forex/ and /markets/capital-markets/, with the signed metadata-only China catalog at /markets/china-macro/. They describe real public coverage and its gaps; they do not claim executable quotes, every security or licensed redistribution rights. China raw exports, histories and values remain restricted and never enter a gauge or score; see docs/NBS_SIGNED_EXPORT.md for the operator and verification contract.

Public money-market contracts

• GET /api/money-markets returns the full USD Money Market Desk. Every metric carries its source, as-of date, cadence and plain-language explanation; meaningful cards add native-series changes, a one-year robust z-score and a three-year empirical percentile. Derived spreads use exact common observation dates—there is no forward-fill across unrelated source clocks. The headline displays both the raw most-extreme channel and a conservative Bonferroni- adjusted desk-wide rank; reserve and ON-RRP stock levels remain visible context but cannot manufacture a funding-stress label by themselves.
• GET /api/v2/money-markets returns the Global Money Market Atlas from already collected canonical observations for all 11 registered packs. Eleven is a catalog count, not a claim of 11 live benchmarks or validated markets. It also returns a 52-row source-audited expansion ledger with benchmark taxonomy, authority link, access/rights caveat, confidence, review stage and verification date. Those rows are discovery metadata, not quotes. Query coverage, faults and per-market status for what is actually available now; the request never starts collection.

Atlas state is explicit. AVAILABLE means a redistributable raw observation is present. DERIVED_CONTEXT means a restricted/derived-only input may contribute non-reversible own-history statistics, while its raw level and history remain withheld. POLICY_ONLY means an official policy anchor is visible but no eligible traded benchmark is available; policy is not used as a substitute. A declared unavailable market remains an evidence gap, not a calm reading. Changes and history windows follow each adapter's native cadence—weekly or monthly series are never padded into daily data—and cross-market comparison uses own-history normalization rather than unlike rate levels.

Seiche's AGPL-3.0-or-later license covers the code. It does not override upstream terms: allowed inputs may expose values, derived_only inputs expose bounded derived context, metadata_only inputs can describe availability but cannot enter public calculations, and prohibited values and source metadata are omitted.

Run it

# backend (Python 3.12+)
cd backend
python3 -m venv .venv && .venv/bin/pip install -e .
.venv/bin/uvicorn seiche.api:app --port 8787

# frontend (dev)
cd frontend
npm install
npm run dev          # http://localhost:5173 (proxies /api to :8787)

# or production single-process: npm run build, then uvicorn serves frontend/dist at /

First load is slow (cold fetch of several years of history); everything after is served from the SQLite cache with cadence-aware TTLs.

The operator CLI

seiche pull               # force-refresh, print the index line
seiche brief --save       # this morning's desk note (markdown, archived to data/briefs/)
seiche alert              # evaluate alert rules once (cron/launchd-friendly; exit 2 = fired)
seiche watch -i 1800      # pull + alert on a loop
seiche replay 2019-09-12  # Time Machine in the terminal
seiche backtest           # PROOF summary
seiche ml                 # ML Lab: event probability + honest validation
seiche analogs            # Tide Tables: nearest historical analogs + forward fan
seiche swell              # the funding-stress forward curve, 6 weeks out
seiche physics            # the physics board: floor, modes, determinism, tail law
seiche bathymetry         # the basin floor in detail: potential, spectrum, entropy, first passage
seiche book               # the Book: today's positions + walk-forward P&L verdict
seiche ask "…"            # desk assistant, grounded in the live board
seiche serve              # API + UI
seiche mcp                # serve the board to AI agents over MCP (stdio)
seiche nbs-intake catalog # four code-owned NBS source identities; no values
seiche nbs-intake status  # verify the public-only signed China revision head

For AI agents (MCP)

Seiche is also a Model Context Protocol server — any MCP-capable agent (Claude Code, Codex, your own) can read the live board as tools. Where a data feed hands an agent raw macro numbers, Seiche hands it the conclusion: a regime read, forward event odds, historical analogs, and a historical diagnostic whose status, misses, and eligibility flags stay attached. Stdlib-only, no new dependencies.

claude mcp add seiche -- seiche-mcp          # Claude Code, local (stdio)
SEICHE_MCP_PUBLIC=1 seiche-mcp               # free surface only

Or, zero-install, over HTTP: the same tools are served at /mcp on the API (https://api.seiche.info/mcp). Add the URL and start calling. Eleven tools answer anonymously, no token, no sign-up, no email:

claude mcp add --transport http seiche https://api.seiche.info/mcp
curl https://api.seiche.info/api/gauge

The copy-paste quickstart and live tool runner are at seiche.info/developers.

That is the conclusion, the precedent, the honest record, the freshness, granular USD money-market context, unified world-market context and cross-market transmission context, and it stays free. Five tools want a bearer token because they read gated forecasting, replay, positioning, prose, or LLM engines rather than a published contextual conclusion: funding_stress_forecast, replay_asof, positioning_book, desk_brief and ask_desk. institutional_flows answers anonymously but keeps its method_versions back for the same reason.

Nothing fails at call time over this: tools/list returns exactly the tools the caller can run, so an agent never sees a tool it cannot use. Both surfaces are rate-limited per caller and metered per UTC day (anonymous callers per IP).

Full setup, the tool catalogue, client config, metering, tokens, and the pay-per-call option for the five: docs/MCP.md.

Want the board in your pocket instead of a client config? The Hermes desk-agent kit turns hermes-agent into a Seiche desk agent on Telegram/Discord/Slack: a scheduled morning brief, regime alerts with anti-noise rules, construction-PIT episode replays, and a PROOF-grounded answer to "can I trust this". Walkthrough: docs/HERMES.md.

Where this is heading on the crypto side (stablecoin reserves, tokenized Treasuries, DeFi rates all sit on the market Seiche reads): docs/CRYPTO.md.

Alerts dedupe per state in SQLite, notify via macOS notification and optional webhook (SEICHE_WEBHOOK_URL — Slack/Telegram/ntfy style {"text": ...}). A launchd template lives in ops/com.seiche.watch.plist.

Production deployment

# Existing canonical host: install or refresh the signed release controller
# without activating it, then run its gate-only acceptance cycle.
bash /home/seiche/app/ops/deploy/install-release-poller.sh
SEICHE_CONTROL_GATE_ONLY=1 /usr/local/sbin/seiche-release-poll

Production uses the operator-provisioned /home/seiche/app layout documented in ops/deploy/RELEASE-POLLER.md; this repository intentionally has no unattended first-VPS bootstrap. The old /opt installer is retired because its service and state layout never matched production.

GET /api/health never starts or waits for the full board build. It returns the last completed snapshot's health fields with HTTP 200, or an immediate HTTP 503 with status: warming_or_unavailable while the cache is cold. The production background warmer owns snapshot construction; deployment probes only observe whether it has published a result.

Put a TLS reverse proxy in front. The box already serving another site (e.g. Palimpsest) just adds a vhost — nginx:

server {
    listen 443 ssl http2;
    server_name seiche.example.com;   # certbot --nginx -d seiche.example.com
    location / { proxy_pass http://127.0.0.1:8787; proxy_set_header Host $host; }
}

Signed releases: the on-host release poller is the sole automatic controller. It accepts only the reviewed SSH-signed main tip, tests that exact SHA in isolation, and hands it to the rollback-owning deploy wrapper. The deploy-hetzner workflow is workflow_dispatch-only break-glass recovery and must not run while the poller is active. The API cache lives under /home/seiche/app/backend/data; canonical market state and backup receipts live under /var/lib/seiche and /var/backups/seiche-market. Follow the poller and market backup runbooks rather than copying one directory as if it were the whole track record. LLM keys for the desk assistant/Navigator and Telegram alert credentials belong in a root-owned production environment file, not the retired ops/deploy/seiche.service template. On the host, create /etc/seiche/operator.env with mode 0600, then attach it to the actual seiche-api.service through a non-secret systemd drop-in:

sudo install -d -o root -g root -m 0755 /etc/seiche
sudo test -e /etc/seiche/operator.env || \
  sudo install -o root -g root -m 0600 /dev/null /etc/seiche/operator.env
sudoedit /etc/seiche/operator.env
sudo install -d -o root -g root -m 0755 \
  /etc/systemd/system/seiche-api.service.d
printf '%s\n' '[Service]' \
  'EnvironmentFile=/etc/seiche/operator.env' | \
  sudo tee /etc/systemd/system/seiche-api.service.d/operator-env.conf >/dev/null
sudo systemctl daemon-reload
sudo systemctl restart seiche-api.service

Use systemd KEY=value syntax in the environment file for only the required SEICHE_LLM_*, SEICHE_TELEGRAM_*, or webhook settings. Never commit their values. Confirm permissions with stat, then check systemctl status seiche-api.service and the API health route after restart.

Tuning the editorial voice

backend/seiche/config.py quarantines every judgment call: composite weights, regime thresholds, resonance/turn/tell parameters, alert rules, the episode library, contract DV01s. The math never hides an opinion.

Non-goals

The eleven-tool public evidence surface needs no account and does not depend on paid upstream data; optional licensed or tenant inputs remain explicitly bounded. Five compute-heavy forecast, replay, positioning, prose and LLM tools are account-gated. Seiche does not claim intraday-tick coverage: daily cadence plus operation results is the honest granularity of the public stack. Historical reconstruction uses final/current-vintage inputs and is therefore construction-PIT, not validated-backtest evidence. The engine blocks the historical backtest output unless every required input has an ALFRED/as-published vintage manifest. From v2 onward Seiche accrues a true as-published point-in-time record (/api/pit) and stores immutable observation captures for forward vintage reconstruction. Those forward captures cannot repair vintages that were never retained. Not investment advice.