Undertow MCP | Market liquidity and exit-cost tools

Endpoint: https://api.seiche.info/undertow/mcp (streamable HTTP, no install)

Undertow is the market-liquidity product of LIQUILENS PRIVATE LIMITED, a registered company in India (CIN U62011RJ2026PTC116792). One company, three connected financial-evidence products: LiquiLens investigates institutions, Seiche explains funding conditions, and Undertow examines market exit evidence. Together they support a review of an exposure from institution to funding to liquidity, with each source date, coverage gap and research limit retained.

Built for people and AI agents working internationally. Access is global; coverage is market-specific. Read the company and investor story or inspect the shared company profile.

Start with a question: Open the crypto workbench. Choose a dollar size and click to read the current published venue estimates, snapshot date and method. No account, key, wallet or client installation is needed for this example. The result is a depth-based estimate, not an executable quote. The same page provides Codex, Claude Code, Cursor and VS Code setup.

Try it live: liquilens-undertow.com/developers · API catalog: api.seiche.info/undertow

Undertow exposes estimated exit cost by position size and venue, the concentration of quoted depth, realized depth-collapse episodes, and liquidity tiers across market segments. This MCP 1.13.0 endpoint exposes 24 read-only tools, split into 16 public and 8 subscriber tools, plus 3 guided prompts. Its capability inventory is pinned to liquilens-undertow commit f103a9929a438f92964e7254edd7553c5a4dfe3a, the hosted implementation at deploy/hetzner/undertow-mcp. The stdio discovery server in undertow_mm/mcp_server.py is a separate discovery surface; it is neither this registry listing nor the public stdio adapter provided here.

The liquidity_tiers result includes public measure counts, replay eligibility and the published reason a score is withheld. A PARTIAL tier can remain even when every measure qualifies. Settlement histories and source receipts are available through the separate source-data MCP.

Add it

Claude Code:

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

Claude.ai / ChatGPT / Cursor: add a custom connector or MCP server with the URL above. No key and no wallet for the free surface.

The free agent starter kit connects Undertow exit estimates and venue concentration with Seiche funding context and LiquiLens bank evidence. Follow the Hermes or OpenClaw guide for the shared setup. It uses public tools under fair-use limits; model-provider costs remain separate.

For a recurring research workflow, use the shared infrastructure setup guide and current interface catalog. They distinguish the public research MCPs, the eight-tool Research Desk and optional private runtimes. Private paper assessment is a separate integration; Undertow's public tools do not submit orders. Source observation dates, data rights and eligibility remain explicit even when hosting and backups are healthy.

This repository contains the discovery manifest, documentation and an optional anonymous stdio adapter for the hosted service. The official registry serves io.github.beepboop2025/undertow version 1.13.0.

Local stdio and container installation

The direct hosted URL above remains the simplest connection. For clients that require stdio, clone this repository at a reviewed commit and use Python 3.12+ and uv:

uv sync --locked
uv run --locked undertow-mcp

A Claude Desktop configuration can use uv as its command and ["run", "--directory", "/absolute/path/to/undertow-mcp", "--locked", "undertow-mcp"] as its arguments. The stdio adapter exposes only the 16 anonymous tools and three prompts; subscriber access uses the direct hosted URL instead.

For Docker and Glama's container build:

docker build -t undertow-mcp .
docker run --rm -i undertow-mcp

The root Dockerfile runs as an unprivileged user and starts stdio directly. No API key, bearer token, wallet, port, volume or environment setting is needed. Outbound HTTPS access to the fixed api.seiche.info endpoint is required. Redirects, environment proxies and arbitrary upstream URLs are disabled. Requests are capped at 64 KiB, responses at 2 MiB, and each upstream operation at 10 seconds including queue wait. No automatic retries or response cache can hide outages or spend a quota twice. Upstream version/catalog drift fails closed. Tool results, native isError values and rights refusals are preserved.

A Glama maintainer can configure the root Dockerfile, complete its build test, and publish a Glama release from the listing's admin page. A GitHub commit or release does not create a Glama release. This repository does not claim a grade until Glama has actually rescanned and inspected it. See Glama's release guide.

Hosted protocol compatibility

• 2026-07-28: stateless requests use server/discover, per-request _meta, MCP-Protocol-Version, and mirrored Mcp-Method / Mcp-Name routing headers.
• 2025-11-25, 2025-06-18, and 2025-03-26: retained legacy initialization, tools, prompts, notifications, batching, and ping behavior.
• Discovery identifies all sixteen public and eight subscriber tools. Anonymous tools/list returns only the public inventory; entitlement is checked fresh on every subscriber request.
• resources/list and resources/templates/list return explicit empty catalogs. resources/read returns a not-found error and never invents a resource.

Example

In the snapshot generated at 2026-08-08T15:01:22Z, selling $1,000,000 of BTC at the selected $1,000,000 published size rung cost 2.386 bp on Binance and 13.623 bp on Bitfinex: about $238.60 against $1,362.30. Gemini was the dearest observed venue in that same snapshot at 25.852 bp, or about $2,585.20.

Venue rankings can change during the day, and those differences are not visible in a single consolidated price. Undertow publishes the observation time and venue inputs with the estimate.

Crypto workbench and skills

The browser workbench combines published BTC exit comparisons with funding and portfolio scenarios. The scenario forms run locally in the browser.

The crypto agent kit contains three installable skills, a standard-library Python client, offline scenario calculators and a repeatable brief connecting Undertow, Seiche and optional LiquiLens institution evidence. Exact file digests accompany the download. Nothing runs or installs automatically.

Tools

Prompts

Commodity futures are intentionally absent from that table. Undertow has no licensed point-in-time depth by contract month, venue and session, so executable commodity exit cost is CANNOT_ASSESS_EXECUTABLE_EXIT_COST; open interest or daily volume is never substituted. For aggregate WTI/Henry Hub cash pressure, Cushing and benchmark structure, call Seiche's public oil_funding_context or use /oil in @seiche_desk_bot.

Limitations

• PARTIAL is not calm. A publishable segment needs enough qualifying measures and a passing replay. The response preserves the exact counts and withheld-score reason; qualifying measures alone do not imply a score.
• Exit costs are estimates, interpolated from published quote depth at the 1% and 2% bands. Never a book walk. The snapshot refreshes roughly hourly, so it is a snapshot and not a real-time feed.
• Crypto measures are still accruing, so the board has not earned a crypto stress percentile and you should never quote one from it.
• The sealed record includes misses. Calls are hash-chained and signed before their outcomes are knowable, then scored against the point-in-time board.
• Commodity execution is a declared coverage gap. Ballast is useful upstream context from Seiche, not a depth ladder and not an Undertow exit-cost estimate.

Research and market data, not investment advice.

Subscriber tier

The Agent and Desk tiers unlock the eight subscriber tools. Send /agent to the Telegram bot to mint a bearer token:

{
  "mcpServers": {
    "undertow": {
      "url": "https://api.seiche.info/undertow/mcp",
      "headers": { "Authorization": "Bearer YOUR_TOKEN" }
    }
  }
}

The token proves identity only. Entitlement is re-read from live membership on every call, so access stops when the subscription does rather than when the token expires. Subscriber tools are invisible to an anonymous tools/list, and agent_access_status tells you where you stand.

Only tools/call is metered, reported on X-MCP-Usage-Used, X-MCP-Usage-Limit and X-MCP-Usage-Remaining. GET /undertow/mcp/usage is the self-meter. Hitting a quota returns a normal JSON-RPC result carrying isError and an upgrade pointer, never a dropped connection.

About this repository

This repo is the listing: a README and the two manifests that let directories describe the server accurately. The server itself is hosted at the endpoint above; its source is deploy/hetzner/undertow-mcp in the Undertow product repository and the registry target remains hosted 1.13.0. The adapter forwards the native public schemas and results without computing market values or granting subscriber access. Its own version is 0.2.0; the upstream contract is 1.13.0.

Verification and deployment boundary

The verification workflow validates the exact 40-character releaseCommit in contract.json against the immutable source receipt in source-receipt.json. The receipt binds that commit and contract to SHA-256 digests of the hosted implementation and registry manifest without granting this public repository access to the private product repository. A maintainer creates or updates the receipt only after running the local pinned-core verifier against a clean checkout; that verifier derives the public/subscriber split, prompt inventory, protocol versions, server identity and registry manifest directly from the source and checks both artifact digests. A branch name or current product-repository HEAD is never accepted as release provenance.

A separate scheduled and manually dispatchable smoke makes anonymous, read-only calls to the listed endpoint. It initializes the legacy protocol, exercises modern discovery, checks the public tools, subscriber advertisement, prompts and explicit empty resource catalogs, then calls agent_access_status. It does not use a bearer token or exercise a subscriber tool. Run the same checks locally with:

uv run --locked python -m unittest discover -s tests -v
python3 scripts/verify_core_pin.py --receipt
python3 scripts/verify_core_pin.py --core /path/to/exact/core/checkout
python3 scripts/smoke_live_mcp.py

The source pin is a reviewed contract boundary, not an HTTP deployment receipt. Undertow's release controller records the deployed SHA and append-only receipts on the host, while the public MCP response currently reports its semantic version and capability catalog but no exact Git commit. Therefore a green live smoke proves the published behavior, not that the host runs this exact SHA. Do not repin this listing from a matching version or catalog alone; require the host's exact successful deployment receipt (or an equivalent authenticated SHA attestation) first.

Siblings from the same lab

• Seiche: US dollar funding stress.
• LiquiLens: bank, NBFC and lender failure risk, and whether that stress is reaching firms and households.
• groundcheck: claim grounding and citation verification for general text.
• Palimpsest: live internet-censorship signals.

Human front door: liquilens-undertow.com and the Telegram desk.

License

The original public integration code and documentation in this repository are MIT licensed. See NOTICE.md for scope: this grant does not cover the private Undertow service or third-party market data. Native access controls, evidence limitations and source-rights holds still apply.