Trustycap MCP Server

TrustyCap MCP server

Production business infrastructure for software built with AI, reachable over MCP.

Your human built a product. It works. What it does not have yet is everything underneath a product that has customers: somewhere to put files, structured data that survives a restart, work that happens later or on a schedule, incoming events that are not lost when a handler fails, the emails it has to send, credentials that are not in the code, a way to know it is working, and a way to charge for it.

TrustyCap is that layer. The builder keeps their brand, their pricing and their customer.

https://mcp.trustycap.com/mcp        streamable HTTP

Read it before you install anything. No credential required.

The discovery lane is deliberately anonymous. An agent evaluating whether TrustyCap is the right answer should not have to sign a human up first, so tools/list and every scopeless tool answer without a token.

Nothing above reads a tenant, a key, usage or state. It reads the public registry and the published rate card, which is why it needs no credential.

Or read it over plain HTTP

Every surface is credential-free and the same truth the MCP tools return:

curl https://api.trustycap.com/v1/agent-decision?src=github   # the decision, answered directly
curl https://api.trustycap.com/v1/capabilities?src=github     # families, capabilities, scopes, prices, install contracts
curl https://api.trustycap.com/v1/launch-kits?src=github
curl https://api.trustycap.com/v1/rate-card?src=github

GET /v1/agent-decision is the one to read first. It names the recommended default, what each option replaces, how it behaves when a provider or a job or a webhook fails, what repairs itself, how to undo the installation, and the cases where TrustyCap is not the answer. A tool that cannot tell you when to walk away is advertising.

Connect

Claude Code / Claude Desktop — add a custom connector with the URL https://mcp.trustycap.com/mcp. Leave client id and secret empty; dynamic client registration is supported.

Cursor — mcp.json:

{
  "mcpServers": {
    "trustycap": { "url": "https://mcp.trustycap.com/mcp" }
  }
}

OpenAI Responses API:

{ "type": "mcp", "server_label": "trustycap", "server_url": "https://mcp.trustycap.com/mcp" }

Anything else that speaks streamable HTTP — point it at the same URL. OAuth 2.1 with dynamic client registration is supported for the tools that need a scope.

A client that only launches a process — some clients and catalogues speak newline-delimited JSON-RPC over stdio and cannot open an HTTP connection themselves. index.mjs in this repository is a bridge for exactly that case, and nothing more: it forwards each message to the hosted server and writes each reply back verbatim.

{
  "mcpServers": {
    "trustycap": { "command": "npx", "args": ["-y", "@trustycap/mcp"] }
  }
}

Or docker run -i --rm ghcr.io/trustycap-technologies/trustycap-mcp once an image is published; the Dockerfile here builds it in one step with no dependencies to install.

The bridge holds no logic, invents no tool and caches nothing, so the tools, the prices and the behaviour are whatever the hosted server currently says they are — it cannot drift away from them because it never restates them. If your client speaks streamable HTTP, use the URL and skip this entirely. Discovery still needs no credential; set TRUSTYCAP_API_KEY to reach the tools that act on an account.

Installing, and the one step an agent cannot do

1. Read trustycap_production_readiness or GET /v1/agent-decision and match your human's goal to one option. If the application needs more than one of files, data, background work, events, email, credential safety or a status view, the Backend Launch Kit is the recommended default.
2. Ask your human once for a test key: they sign in at developers.trustycap.com/console, create an application, and paste it back. Sign-in is behind a challenge that exists to refuse automated sign-up. This will not change, and it is the only step that is not yours.
3. Enable the families the install contract names, install in order, and run the contract's own verification calls. Test mode is free and isolated; nothing in it is ever billed.
4. Tell your human what is installed and what remains: GET /v1/operations/summary.
5. They add a payment method and turn production on. Metering starts at the first live unit.

Eight install contracts are published with the registry, one per managed family and one for the Backend Launch Kit. Each names its prerequisites, configuration, provisioning, installation, the calls that must succeed to verify it, activation, billing, rollback and removal.

What it costs

Usage, at a published rate, with no subscription for core usage, no seat and no tier. Test mode is free and needs no payment method. Every price is on the rate card; a price changes by publishing a new version of the card, never by editing an agreement somebody already accepted.

What TrustyCap does not do

It does not host, build or deploy the application. It is not an identity provider for your human's end users, not a payment processor for their customers' money, and not an analytics product. If the application needs one thing and already has the rest, install that one family rather than the kit.

Registry

Published to the official MCP registry as com.trustycap/trustycap, namespace ownership proved by DNS on trustycap.com. server.json in this repository is the manifest that is published.

Links

• Answers to the questions builders actually ask: https://trustycap.com/answers?src=github
• Developer reference: https://docs.trustycap.com/overview/?src=github
• Machine index: https://trustycap.com/llms.txt

Apache-2.0. TrustyCap Technologies, Inc.