Metadata-Version: 2.5
Name: roost-mcp
Version: 0.1.0
Summary: Roost public MCP servers for AI agents: a payment pre-flight check (roost-corpus-mcp) and a finance-agent safety test (roost-eval-mcp). Advisory, zero-config, no API key.
Project-URL: Homepage, https://roostcapital.ai
Project-URL: Repository, https://github.com/yahyabayian/Roost
Author-email: Roost <developers@roostcapital.ai>
License: Apache-2.0
Keywords: ai-agents,evals,finance,mcp,model-context-protocol,payments,roost
Classifier: Development Status :: 4 - Beta
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Requires-Python: >=3.11
Requires-Dist: httpx>=0.27
Requires-Dist: mcp<2,>=1.2.0
Description-Content-Type: text/markdown

# roost-mcp

Two zero-config, advisory MCP servers from [Roost](https://roostcapital.ai) for AI agents that handle money. No API key, no signup — an anonymous per-install token is created on first use. Calls are anonymized (hashed identifiers, bucketed amounts, no PII) and aggregated into a benchmark of money-moving-agent behavior; that's why they're free.

Install:

```
pip install roost-mcp
```

This provides two commands:

## `roost-corpus-mcp` — payment pre-flight check

A second opinion before your agent sends a payment. The `check_payment` tool takes an invoice, purchase order, or transfer and returns a clear recommendation — **proceed**, **hold & review**, or **consider financing** — with the reasons: a suspected duplicate, an amount out of line with this agent's own history, an early-pay discount worth capturing, or terms worth using.

Advisory, not a gate — your agent stays in control. Fast deterministic checks, **not** fraud detection or a guarantee; the anomaly checks sharpen as your agent builds a short history. USD invoices, purchase orders, and transfers today.

## `roost-eval-mcp` — finance-agent safety test

Find the financial decisions your agent gets dangerously wrong before it touches money. `list_eval_scenarios` returns money-action situations (fraud bait, payee changes, overspend, duplicate/replay, social engineering, …); your agent decides proceed/hold/flag on each; `submit_eval` returns a private scorecard that leads with a machine-readable headline — `summary.status` (`balanced` / `unsafe` / `over_conservative` / `mixed` / `no_data`), `assurance_level="self_reported_unverified"`, and a verbatim `disclaimer` — followed by the component metrics (coverage, correctness, consistency, severity-adjusted risk, unsafe-proceed and conservative-friction rates) and the failed cases and fixes. There is no single "safety score" headline.

This is a private, self-reported, unverified self-test — do not publish or badge an anonymous-tier result; verified, trace-backed evaluation is a separate tier.

It is an **advisory self-test, not a certificate**: a determined caller can harvest answers across runs, so results are a private diagnostic, not a third-party guarantee. Verified evaluation (trace-backed / Roost-driven) is a separate, future tier. Do not submit personal, payment, or confidential business information in free-text fields.

## Configuration

- `ROOST_CORPUS_API_BASE` — API base (default `https://roost-corpus.fly.dev`, the live public endpoint).
- `ROOST_CORPUS_TOKEN` / `ROOST_CORPUS_TOKEN_FILE` — anonymous token override / file (default `~/.roost/anon_token`).
- `ROOST_CORPUS_SOURCE` — optional install-attribution tag.

Apache-2.0.
