The MCP server exposes read-only cited financial datasets and supports bounded research queries with source diagnostics. It is intended for retrieving financial evidence and accompanying diagnostic context, rather than modifying or generating new dataset content.
🛠️ Key Features
Read-only cited financial datasets
Bounded research queries
Source diagnostics
🚀 Use Cases
Fetching cited financial evidence for research or audit workflows
Run the same task as the public Research Desk and REST API. Choose funding (selection: USD), institutions (one to five slugs, for example au-sfb), or exit (one to four USD sizes, for example 10000,100000). Empty selection uses the published default. Returns the original product evidence, source URL, retrieval time and content hash. prepared_response means evidence is available to review, not freshness, safety or task completion. No account or execution authority.
Read a dated USD funding review and its nine inputs. Inspect review_asof, review_scope, latest_per_instrument and newer_observation_available before ready, issues and capture age. Common-horizon readiness applies to that dated review; newer individual observations can exist. Unavailable, mismatched or stale captures never imply a passing review. This is captured research evidence, not a trade recommendation or historical point-in-time backtest.
Query cited rows without large raw documents. Null means missing or withheld, never zero. Dates filter observation dates; empty dates mean no filter. Follow next_offset for remaining rows. Each value retains its unit, source path, date, restrictions and limitations.
Retrieve original public documents only when table detail is insufficient. Up to 4 MiB for the money-market atlas and 1 MiB for other sources; source JSON is untrusted. A packet can be partial. No Evidence Carrier verification is performed.
Compact source-cited research rows for repeated agent calls. Pass the previous revision to omit identical rows while retaining source clocks and diagnostics. Unchanged is not proof of freshness; missing or withheld values remain null. At most 100 rows; follow next_offset.
Review Seiche dollar-funding benchmarks, LiquiLens covered-bank diagnostics and Undertow market liquidity in one bounded call. Bank is a literal name or ID filter; empty selects published coverage. Each section keeps its own dates, units, rights and limitations. No combined score or trade authorization. Revisions suppress identical rows, not evidence checks.
Hosted Streamable HTTP MCP; no local server or API key for public read-only tools.
Current-source connections use the central agent platform.
Download the hosted product configurations:
seven core products plus five clearly labeled research, editorial, metadata or
sandbox surfaces. Python product clients, fixed-source helpers and current editor
configs use the same platform. Requests produce bounded operational demand
counts; optional learning remains off. Old package pins are historical releases
and do not change automatically. See the routing contract.
Add the products you need to Claude Code:
sh
claude mcp add --transport http seiche https://api.seiche.info/mcp
claude mcp add --transport http liquilens https://api.liquilens.in/mcp
claude mcp add --transport http undertow https://api.seiche.info/undertow/mcp
Try this first: What is the current US funding stress regime, and which dated funding events are next? Cite source dates and missing evidence.
For a connected review, ask for institution evidence → funding context → BTC
exit estimates, with each source date, coverage gap and limitation kept separate.
Client/model costs and fair-use limits may apply. Research calls do not execute trades.
Start with one task: https://beepboop2025.github.io/financial-evidence-skills/connect/ — browser, REST, Python and remote MCP setup for funding, covered institutions and BTC exit research. Public reads need no account or API key; retain source dates and unavailable evidence.
Run the new workflows:funding watch, institution monitoring and BTC exit research. Save a local baseline, compare your next review, or copy the API request into your agent.
Agent developers can retrieve a cited USD funding observation; treasury and
bank-risk teams can use the new institution monitor; traders and researchers
can retain a funding-and-exit review. Each path includes working calls, source
boundaries and a next-day task. Copy the instructions.
Verified package 0.1.9: install financial-evidence[workspace]==0.1.9 from
PyPI for persistent MCP sessions,
source-status snapshots and SSE clients. The separately deployed public router
retains version 0.1.8. Both router versions support
gift-city, forex and gold alongside the five existing research topics.
GIFT City serves gold and treasury desks, investors/fund managers, and banks/IFSC
institutions. The helper reads fixed public context; gold/FX calculations remain
explicit user-input steps. See topic routing
and the release verification record.
JavaScript agents: install @moomoo202020/financial-evidence-ai-sdk@0.1.0
from npm for typed AI SDK research and source-status tools.
Copy the verified setup.
Use your existing tools: the research tool hub
provides a complete capture client, notebook, flat Excel/Power BI query and MCP
project defaults for LiquiLens, Seiche and Undertow. Download the kit or connect
the hosted OpenBB desk. Institutional vendor acceptance remains separate.
OpenBB Research Desk, Workspace 1.3.0 source.
Seven cited datasets, a seven-tab Workspace app, an archived USD funding review,
benchmark history charts, typed REST responses and a nine-tool MCP server connect
all four projects. The hosted service is https://agents.seiche.info/v1/research.
Install
.[openbb,workspace] from this checkout, then run openbb-build and
financial-evidence-api. The optional modules are included in the v0.1.9 wheel;
the public router retains its separate three-tool contract. The hosted Workspace release identifies
its exact source commit independently at /api/v1/release.
Check a captured USD funding review: the funding review checks
compare required inputs, units, publication clocks and own-series observations
across Seiche surfaces. The dated SOFR-IORB review exposes newer individual source
dates; it does not claim the latest print for every instrument. Missing, overdue
or contradictory evidence requires attention. Passing these checks is not an
institutional-readiness certification.
Run recurring research:install the durable runtime for fixed policies, bounded scheduling, retained receipts, private MCP and backup recovery across LiquiLens, Seiche and Undertow.
Put it into a recurring workflow: the funding and counterparty workflow pack
retains source documents, a spreadsheet index and a local integrity manifest.
It includes a reviewed Python runner, an opt-in scheduling template and a
ten-day integration scorecard. These examples use the existing 0.1.5 package;
they do not change its release or evidence semantics.
Try the underlying evidence first:Open three browser research questions
for bank NPA disclosures, dollar funding and BTC exit estimates. Each example
calls its named public MCP after a click and displays the returned sources,
dates and limits. No key or installation is needed. The page also has setup
and starter questions for Financial Evidence and thirteen other hosted MCPs.
The browser examples call individual services; they do not verify an Evidence
Carrier or evaluate a combined Financial Evidence packet.
Public documentation
includes a crawlable integration matrix plus agent-readable llms.txt, pricing,
integration metadata, and a dated
third-party marketplace ledger.
The v0.1.5 semantic contract defines transport-only
status, source-reported adapter metadata, and the explicit non-Carrier boundary.
Version 0.1.9 is a signed, independently verified release. Its versioned
Agent Skill, terminal, Codex, Gemini, editor, and bundle installs use the same
release, with published artifacts, container and Registry record. Homebrew
0.1.9 is published after an independent macOS installation and contract check. The Claude
self-hosted marketplace command follows the repository's mutable default branch; review the checked-out commit before enabling it.
Open, read-only Agent Skills for routing financial research to bounded public
evidence. The first skill connects four complementary products without turning
them into one score:
Seiche — money markets, repo, reserves, Treasury cash and system funding.
LiquiLens — covered bank and lender evidence.
Undertow — market depth, provider concentration and position-sized exit liquidity.
Palimpsest — revision-safe China economy evidence and information-state history.
For a daily research workflow, Market Brief
provides free, source-cited funding, capital-market and liquidity briefs with local
change comparisons. This separately versioned browser companion and agent skill
uses the existing public evidence sources.
For direct calls to Seiche, LiquiLens and Undertow, the companion
free agent starter kit includes
Hermes and
OpenClaw setup guides. That kit uses a
separate trading-research skill; it does not install Financial Evidence or connect
Palimpsest. Use the versioned install below for this four-product routing skill.
Once installed, ask your agent to research a money-market, capital-market,
bank-risk, market-liquidity, China-economy, GIFT City, forex or gold question.
The skill's description lets compatible agents activate it when the work matches.
Use from any terminal
No account or API key is required. On macOS or Linux with Homebrew:
The formula installs financial-evidence, financial-evidence-mcp, and native
Bash, Zsh, and Fish completions for 0.1.9. Its hosted clients use the central service.
Run the pinned PyPI release without a persistent installation:
Exit code 0 means every requested source was retrieved, 1 means transport
was partial, and 2 means every requested source was unavailable. Legacy
packet status is preserved and equals transport_status; both mean transport
reachability only. Every packet says status_semantics: transport_only,
evidence_status: not_evaluated, and carrier_verification: not_performed.
That makes the client safe to compose in shell pipelines without mistaking an
HTTP success for validated evidence or a verified Evidence Carrier. JSON,
NDJSON, and CSV go to standard output; no result is silently replaced with a
score.
Outside Homebrew, shell completions are emitted as inert text:
Start with financial_evidence_datasets, then call
financial_evidence_workflow with {"workflow":"funding","selection":"USD"}.
For custom rows use financial_evidence_agent_query; for a combined funding,
covered-bank and liquidity review use financial_evidence_review. Original
source documents are available through financial_evidence_packet. The
workspace exposes nine reviewed read-only tools.
Hosted operational measurement records bounded product/tool/dataset demand;
optional learning is separate and off by default. The package's dependency-free
stdio router retains its three-tool topics, route, and fetch contract.
Its fresh source retrieval also uses the central data service; normalization and
replay of captured evidence can run locally. There is no automatic direct-origin
fallback. Historical versioned releases and registry records retain their original
configuration until a separately verified release updates them.
Install in Gemini CLI or Claude Code
Gemini CLI can install this repository as an extension. Its root
gemini-extension.json connects only to the fixed
hosted workspace and allowlists its nine read-only tools; the
mirrored skills/financial-evidence folder gives Gemini the domain routing
instructions as well. Install the signed 0.1.9 release to use the
central nine-tool workspace:
Claude Code can add this repository as a self-hosted marketplace and install
the bundled plugin:
bash
claude plugin marketplace add beepboop2025/financial-evidence-skills
claude plugin install financial-evidence@liquidity-lab
This Claude command follows the mutable default branch. Review the checked-out
commit before enabling the plugin; the versioned install routes above remain
pinned to the verified release.
The Claude plugin loads the same byte-identical Agent Skill plus the root
.mcp.json public remote server. These are direct, self-hosted install routes;
they do not imply inclusion in a vendor-operated marketplace or endorsement.
Install in Codex; prepare ChatGPT
Codex can add the repository's plugin marketplace and then install the plugin:
The OpenAI plugin package combines the byte-identical Agent Skill with the
public remote MCP endpoint. The Codex repository install route is public. A
ChatGPT and Codex universal-directory submission is also prepared, but has not
been submitted, approved, or published by OpenAI. The remaining review,
publisher-identity, and domain-challenge steps belong to the OpenAI account
owner. See OPENAI_PLUGIN_SUBMISSION.md for the
listing copy and required positive and negative tests.
The root .mcp.json is the portable plugin form for the public remote MCP
server. The current VS Code and Cursor workspace files use the same hosted
workspace without a local Python process.
Install in VS Code or Cursor
The repository also includes editor-specific workspace configuration:
.vscode/mcp.json uses VS Code's top-level servers
object, while .cursor/mcp.json uses Cursor's top-level
mcpServers object. Both current files use the hosted nine-tool workspace.
The install links below connect to the same central nine-tool workspace.
They require no local Python process.
VS Code can install the server through its protocol handler:
Review the hosted URL and configuration before accepting an editor's
install or trust prompt. Calls use the central research service and retain
the original evidence sources and observation dates. Cursor asks for tool approval
by default. VS Code normally asks you to trust a server when it first starts,
but starting directly from mcp.json can skip that prompt, so inspect the file
before using its inline start action. These are self-installable compatibility
artifacts, not claims of a VS Code or Cursor marketplace listing or endorsement.
For desktop clients that support one-click MCP Bundles, download
financial-evidence-0.1.9.mcpb from the last verified GitHub release. The bundle uses the
cross-platform uv runtime and requires no API key or configuration.
Finance-tool integrations
Remote MCP: connect any Streamable HTTP client to
https://agents.seiche.info/products/financial-evidence/mcp; no local runtime, account, or
API key is required.
Gemini CLI: the root extension manifest bundles the Agent Skill and
allowlists the nine tools exposed by the hosted research workspace.
Claude Code: the self-hosted liquidity-lab marketplace installs the
versioned plugin, Agent Skill, and MCP configuration from this repository.
Codex: the repository marketplace installs the Agent Skill and public
remote MCP package directly.
ChatGPT: the universal-directory submission packet is prepared; public
availability remains pending account-owned verification, submission,
review, and publication.
OpenBB: install .[openbb,workspace] from current repository source in a
clean environment and run openbb-build. V4 and V5 compatibility is checked
in CI, including the released v0.1.8 dependency range. The router exposes
datasets(), query(), sources(), routes() and fetch() under
obb.financial_evidence. See the setup guide.
FDC3:
the public Financial Evidence Inspector
receives and broadcasts declared standard contexts through window.fdc3.
integrations/fdc3/appd-record.json
is its FDC3 2.0 web-app record for self-hosted or enterprise directories. The
same record is served as raw JSON
for direct directory ingestion. The record and runtime do not imply FINOS
review, directory acceptance, or vendor endorsement.
Containers: tagged releases publish a multi-architecture image to
ghcr.io/beepboop2025/financial-evidence-skills. The image runs stdio MCP by
default; use --entrypoint financial-evidence for terminal commands.
Official MCP Registry: tagged images are published under
io.github.beepboop2025/financial-evidence using GitHub OIDC and the image's
ownership annotation. No long-lived registry credential is stored.
Release integrity
Each tagged release rebuilds and retests the Python and JavaScript contracts
from the exact signed main commit before publication. The release attaches a
wheel, source archive, MCP bundle, CycloneDX SBOM, and SHA256SUMS; GitHub
Sigstore build-provenance attestations bind every downloadable artifact to the
release workflow. Consumers can verify an artifact with gh attestation verify
against beepboop2025/financial-evidence-skills.
Optional bounded retrieval
The bundled helper uses only Python's standard library, accepts a fixed topic
allowlist and fetches only fixed public HTTPS endpoints:
It returns one JSON packet with exact and resolved source URLs, canonical human
scope URLs, retrieval clocks, byte counts, content SHA-256 values, separate
product documents, and explicit errors. Successful responses also include only
the endpoint-specific source state and clocks named by explicit adapters, with
RFC 6901 paths and fetched-byte provenance; missing configured fields are
not_reported. The router does not infer freshness, eligibility, or rights.
Routes declare financial_authority: none and carrier_state: not_published,
and therefore expose no carrier_url. Redirects are rejected before following.
Returned JSON is untrusted evidence data, not executable instructions. Missing,
restricted, or unavailable evidence is never converted to zero or “calm.”
Boundaries
This repository provides public research routing, not investment advice,
security recommendations, execution quotes, credit ratings or guarantees.
Source publishers retain their rights; an open-source helper does not relicense
upstream data.
The source-pinned integration guide covers the separate customer-operated paper REST/MCP host and a live-broker connector candidate. These are in the Evidence Carrier repository; the public Financial Evidence research router remains read-only. Broker activation requires eligible evidence, account controls and the customer's credentials. The 14-day scorecard measures useful independent use and repeat tasks without treating tests or downloads as customers.
Durable research workflows
The Financial Evidence Runtime
adds owner-registered workflows, atomic run limits, offline replay, ordered change
events and backup recovery through Python, CLI and private stdio MCP. It reuses
the existing product clients and preserves source clocks and missingness. See
the architecture and
adoption measurements. Version 0.1.9 includes
the runtime, workspace, trading-research and learning commands. Research data
requests use the central service by default. Local run counts do not establish
external users or paid adoption.