CSPR Sentinel
An MCP server for autonomous AI agents to discover, pay for, and rate paid APIs on Casper with policy-controlled approvals and on-chain reputation.
README
CSPR Sentinel
CSPR Sentinel is a policy-controlled payment and reputation layer for autonomous AI agents on Casper. It exposes an MCP server that lets agents discover paid services, request x402 purchases, wait for human approval when required, and build bilateral reputation from verifiable outcomes.
- Live demo: cspr-sentinel.vercel.app
- MCP endpoint:
https://cspr-sentinel.vercel.app/api/mcp - Facilitator discovery: embedded
/supportedendpoint - Source: github.com/yi-dong-z/cspr-sentinel
The included RWA diligence agent demonstrates the full flow with synthetic commercial paper data:
- It buys a low-cost asset verification service automatically.
- It sends a higher-cost risk intelligence request to the owner approval queue.
- After approval, the request is paid and delivered.
- The application stores the Casper transaction hash and anchors reputation aggregates on-chain.
The provider datasets are synthetic and are not investment, legal, or compliance advice.
Why it exists
Agents need the ability to pay for one API call at a time, but unrestricted signing keys are unsafe. CSPR Sentinel separates planning from authority. The model can request a purchase; deterministic policy code decides whether it is denied, executed, or held for owner approval.
Architecture
flowchart LR
A[RWA Agent] -->|MCP tool call| M[CSPR Sentinel MCP]
M --> P[Policy engine]
P -->|small request| X[x402 client]
P -->|large request| H[Owner approval]
H --> X
X --> S[Paid API]
S --> F[Casper facilitator]
F --> C[Casper Testnet]
C --> L[Payment evidence]
L --> R[Reputation registry]
P --> N[(Neon)]
R --> N
Monorepo
apps/web: Next.js dashboard, API routes, paid provider endpoints, x402 and contract adapters.packages/core: policy engine, purchase state machine, reputation formulas, Neon repository.packages/mcp: five MCP tools over Streamable HTTP.packages/agent: Anthropic RWA planner and deterministic fallback.contracts/reputation: Odra contract for counters, payment receipts, and rating uniqueness.
MCP tools
| Tool | Purpose |
|---|---|
list_services |
Filter paid APIs by category, price, and provider reputation. |
request_purchase |
Run policy checks and auto-pay or create an approval. |
get_purchase_status |
Read policy, approval, payment, delivery, and evidence state. |
submit_provider_rating |
Submit one verified-buyer rating per delivered purchase. |
get_reputation |
Read agent or provider score, counters, label, and chain anchor. |
Purchase states are requested, policy_denied, pending_approval, approved, paying, settled, delivered, and failed.
Local development
Requirements: Node.js 20 or newer and pnpm 10.
cp .env.example apps/web/.env.local
pnpm install
pnpm dev
Open http://localhost:3000. The default DEMO_MODE=true uses simulated deploy hashes and clearly labels the interface as Simulation mode. Anthropic is optional in this mode; without a key, a deterministic planner creates the same two-service diligence plan.
Run verification:
pnpm typecheck
pnpm test
pnpm security:scan
pnpm build
Neon persistence
Set DATABASE_URL, apply the included Drizzle migration, then start the app:
pnpm db:migrate
pnpm dev
Without DATABASE_URL, the app uses a process-local memory repository suitable for the interactive demo and tests.
GitHub owner access
Set GITHUB_ID, GITHUB_SECRET, NEXTAUTH_SECRET, and OWNER_GITHUB_LOGIN. Only the allow-listed GitHub login can create approval decisions in production. A matching x-demo-admin-key header is also accepted for controlled API automation. MCP, direct purchase, and rating requests use the separate x-agent-api-key header.
Casper Testnet mode
Follow docs/TESTNET.md. Real mode requires a funded agent wallet, WCSPR, the official Casper x402 facilitator, two provider payees, and the deployed reputation contract.
When all variables are set, change DEMO_MODE=false. The runtime then:
- creates EIP-712 payment authorizations with
@make-software/casper-x402; - settles through the configured facilitator;
- stores the returned Casper transaction hash;
- calls
record_purchaseandrecord_provider_ratingon the reputation contract.
Private keys are read only from server environment variables. They are never returned by an API, stored in Neon, or shipped to the browser.
Before starting real mode, validate every required setting and probe the Casper RPC and facilitator:
pnpm testnet:preflight
Real mode fails closed: an incomplete x402 or reputation configuration will stop the request instead of silently producing simulated transaction hashes.
Deployment
The public deployment runs on Vercel with Neon PostgreSQL persistence. It intentionally remains in clearly labelled simulation mode until the generated testnet wallets receive CSPR/WCSPR and the reputation contract is deployed; DEMO_MODE=false is the explicit cutover switch.
- Create a Neon database and apply the migration.
- Import this repository into Vercel with root directory
apps/web. - Add the environment variables from
.env.example. - Set
NEXT_PUBLIC_APP_URLto the deployed origin. - Add the production OAuth callback URL to the GitHub OAuth app.
- Verify
/api/state,/api/mcp, one auto-payment, and one approval payment.
Vercel and external monitors can use /api/health. It returns HTTP 503 when real mode is incomplete, without exposing secret values. GitHub Actions verifies the TypeScript application, security scan, Odra tests, and a reproducible optimized contract WASM build.
Reputation model
- Provider: 60% verified buyer rating and 40% delivery success.
- Agent: 50% payment success, 30% policy compliance, and 20% approval acceptance.
- New subjects start at
50 / Unproven. - Crossing an automatic threshold is compliant; exceeding a hard allow-list, quote, reputation, or daily-budget rule is not.
- Score weights live in the application. Raw aggregates and payment evidence live in the contract so weights can evolve without migrating history.
References
Licensed under Apache-2.0.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.