Their rules.

By Andrés Peña Mellado

Stellar released Raven, an MCP server that gives any AI agent its official documentation and live network data. Installing it in Claude Code or Codex takes two commands. What changes is who can start building on Soroban.

Building on a new blockchain was never hard because of the language. It was hard because of the context. The documentation lives in one place and the standards in another. Network data changes every day. And the AI agent you have open knows none of that: it writes plausible code for a chain it does not know.

Stellar decided to solve that from within. In its Building with AI section, it published three tools to help an agent build on Stellar and Soroban, its smart contract platform. The main one is Raven, which the foundation describes as the recommended path.

The practical result is this: someone with Claude Code or Codex installed gives the agent a URL and signs in once. From then on, the agent searches the official documentation and queries live ecosystem data before writing a line of code. Raven removes the barrier that remained.

Raven does not deploy a contract to mainnet or decide whether an idea is any good. It provides context. Judgment still belongs to the person building.

What is Raven, and why does it matter that it is an MCP?

Raven is a remote MCP server hosted at raven.stellar.buzz. MCP stands for Model Context Protocol, an open standard Anthropic published in late 2024 so a language model can connect to external tools in a uniform way. If an agent speaks MCP, it can use any MCP server. That is why Raven works in Claude Code, Codex, Cursor, VS Code, and Claude Desktop.

The server exposes two tools. search finds Stellar documentation and ecosystem information. execute runs queries against live network data. Underneath, according to its own documentation, Raven wraps Stellar Docs (indexed with Algolia), Lumenloop, Stellar Light, and ecosystem skills. All of that returns as one response.

Raven crosses several sources and summarizes them. That is useful, but it also means its answer comes with its own framing. What a builder verifies is the documentation page or blockchain data Raven points to. The answer saves the search; it does not replace the source.

The code is public. The repository is at stellar-experimental/stellar-raven, under the Apache-2.0 license, with Tyler van der Hoeven (@kalepail) as its principal author. It describes the service as a single authenticated gateway to official documentation, live ecosystem data, community intelligence, and proven playbooks.

Tellus Cooperative has already published an English overview of Raven that sums it up this way: the problem is not that information does not exist; it is that the information is scattered.

It runs on Cloudflare Workers and, according to that repository, has been in production since July 2, 2026. The organization's name tells you what it is: experimental, even though the service is already in production.

How do you install it in Claude Code?

One command and one authentication step. In the terminal:

claude mcp add --transport http stellar-raven "https://raven.stellar.buzz/mcp"

Then, inside Claude Code, type /mcp, choose Authenticate, and complete the sign-in in your browser. Raven uses OAuth. That login enables both tools; without it, the server appears connected but does not respond.

From there the agent can use search and execute on its own. One way to check is to ask it something it could know only by consulting the documentation. For example, ask which SEP governs a deposit and withdrawal flow with an anchor, and see whether it cites the exact page.

When I tried it, the answer came back with four results from the official documentation, each with its URL and section: SEP-6 for programmatic deposits and withdrawals, SEP-24 for those hosted by the anchor, and the difference between them (with SEP-24, the client does not collect KYC information).

For what to try next, Tellus's overview offers a list of starting questions: how Horizon and RPC differ, what documentation to read before issuing an asset, and which ecosystem projects work on wallets and payments. They make a first test for any installation.

And in Codex?

Two commands, the second for login:

codex mcp add stellar-raven --url "https://raven.stellar.buzz/mcp" codex mcp login stellar-raven

The rest is the same. For Cursor, the docs provide a block for ~/.cursor/mcp.json, and for VS Code a code --add-mcp command with the same URL. In Claude Desktop, go to Settings, Connectors, add a custom connector, and paste the endpoint. Anyone using another client has the mcp-remote bridge.

There is also a playground in the browser for trying Raven without installing anything. It asks you to sign in.

The other two tools: skills and llms.txt

Raven is the gateway. Stellar Skills are the manual the agent reads once it is inside: packaged instructions that tell it how to do each thing on Stellar.

The foundation organizes them into seven areas: Soroban contracts; frontend and wallets; assets and SAC; the RPC and Horizon APIs; payments between agents; ZK testing; and the ecosystem's SEP and CAP standards.

There is also a community installer, stellar-build, that adds 46 skills with one command. It comes from the community, not the foundation, and it is worth reading the script before running it, as with any curl | bash.

The third piece is an llms.txt file at developers.stellar.org/llms.txt. It is a simple format: documentation summarized and organized for a model to read as context. It helps when the agent does not support MCP or when you want to give it the entire map at once.

An entry point for new builders

Until now, starting on Stellar meant reading a lot before writing a little. A Soroban contract is written in Rust, standards have numbers, and no model holds the network's data in its head. Raven puts all three within reach of the agent you already had open.

What it does not provide is the idea, the decision about what to deploy, or the review that confirms a contract does what it says. An agent with good documentation writes better code; it does not replace the person reading it.

And Raven can fail: answer incompletely, mix contexts, or cite something outdated. The project acknowledges this and asks that each failure be filed as an issue in the repository.

Tellus's overview suggests a practical way to do that: ask the same agent to draft the issue with the question, its answer, what was missing, and the correct sources. A failure reported that way improves the tool for the next builder.

Raven is also a remote server that requires a login: every query passes through Stellar's infrastructure. The repository says execute queries run in isolated sandboxes without network access. That is something to keep in mind before connecting, not a reason to avoid it.

With the context barrier lower, the coming months will show how many people in Chile move from watching Stellar to deploying something. That number, rather than the tool itself, will tell us whether Raven worked.