Metis RWA Research

Metis MCP

An MCP server for researching tokenized real-world assets on Solana.

It gives an assistant twelve tools: six that read chain and market state live, and six that reach the research the Metis agent has already published. Every figure it returns comes from a fresh read, stamped with its source, the time it was taken, and a hash of the raw payload.

The point is the first question anyone should ask about a tokenized asset and almost nobody can answer quickly: who controls this token, and what happens to my position if they use that control?

Three ways to connect

As a plugin. The easiest path, and the only one that also carries the skills. In Claude Code:

/plugin marketplace add metis-rwa/mcp

/plugin install metis@metis-rwa

In Claude Desktop or on the web, open Customize, then Plugins, then add from a repository with https://github.com/metis-rwa/mcp.git. The plugin connects to the hosted endpoint and adds /metis:due-diligence, /metis:premium-watch, and a token-safety skill that runs whenever someone asks who controls a token. See plugin/README.md.

Hosted, nothing to install. The same tools are served over HTTP at https://metisagent.co/mcp:

claude mcp add --transport http metis https://metisagent.co/mcp

The hosted endpoint answers observation questions from the research agent's own stored history, which goes back further than any single machine can. It reads only, needs no key, and holds no session.

Local, over stdio. Run the package when you want your own RPC endpoint, your own asset list, or history kept on your machine:

npx -y @metisagent/mcp

Or clone and build:

git clone https://github.com/metis-rwa/mcp.gitcd mcpnpm installnpm run buildnode dist/index.js

The server speaks MCP over stdio.

Connect the local server

Claude Code:

claude mcp add metis -- npx -y @metisagent/mcp

Claude Desktop, in claude_desktop_config.json:

{  "mcpServers": {    "metis": {      "command": "npx",      "args": ["-y", "@metisagent/mcp"],      "env": {        "SOLANA_RPC_URL": "https://your-endpoint.example/rpc"      }    }  }}

Any other MCP client works the same way: run metis-mcp (or node dist/index.js) and speak MCP on stdin and stdout.

Tools

On-chain checks:

Research:

Every tool takes either a registry symbol (TSLAx) or a raw Solana mint address, so assets outside the registry work too.

The registry ships with the 72 xStocks on Solana, from AAPLx through XOMx, including the index and commodity fund shares (SPYx, QQQx, GLDx, PPLTx, TBLLx) and the private-market listings (SPCXx, VCXx). Filter it with list_rwa_assets, by substring or by asset type.

Two prompts ship with the server: rwa_due_diligence walks an asset through controls, supply, concentration, depth, and premium, and premium_watch reads a price gap against stored history. Two resources are exposed as well: metis://registry and metis://methodology.

What get_token_controls is for

A tokenized equity is a claim on something held off chain, so the issuer keeps powers a plain token does not have. Token-2022 makes those powers explicit on the mint, and this tool translates each one into what it means for a holder:

• Permanent delegate: tokens can be moved or burned out of any account without the holder signing.
• Freeze authority: any account can be frozen, which blocks selling and redemption.
• Transfer hook: every transfer runs through a program that can reject it, which is how an allowlist is enforced.
• Default account state frozen: new accounts cannot transact until the issuer thaws them, so the token is permissioned by default.
• Pausable: transfers can be halted chain-wide.
• Scaled UI amount: displayed balances are rescaled, which is how splits and similar corporate actions land on chain.

None of these are defects. Redeemable backed assets need most of them. They are simply facts about the asset that belong in any answer about whether it is worth its reference price.

Configuration

Set SOLANA_RPC_URL to an endpoint you control if you can. Public nodes throttle the heavier reads, and get_holder_concentration is the first one they refuse.

METIS_ASSETS_FILE takes an array. Only symbol and mint are required:

[  {    "symbol": "AAPLx",    "name": "Apple xStock",    "mint": "<mint address>",    "decimals": 8,    "assetType": "tokenized_equity",    "referenceTicker": "AAPL",    "issuerName": "Backed Assets"  }]

Verify a mint against the chain before adding it. A wrong address returns confident, wrong answers. Everything already in the registry was read back from Solana first: the names and symbols in src/assets.ts are the ones written into each mint's own Token-2022 metadata, not labels copied from a token list. Check them again at any time:

npm run verify-assets

How an observation cycle works

observe_asset reads supply and holder concentration from Solana, pooled liquidity and volume from DEX venues, and the token and reference prices from a public price service. It compares the result against the observations already stored for that asset, then reports:

• Observations: what is true right now, each tied to the reads behind it.
• Detections: metrics that crossed a threshold, with the baseline they crossed it against.
• Conflicts: sources that disagree, recorded rather than averaged away.
• Confidence: computed from source coverage, source failures, cross-source agreement, and how much history backs the comparison. A model's own stated confidence is never an input.

History lives on the machine running the server. The first call on an asset has no baseline, so only the premium threshold can fire. Call it again over time and supply, liquidity, activity, and concentration detections come alive. The hosted endpoint skips that warm-up entirely: it compares against the history the research agent has been recording every ten minutes since it went live.

Limits worth stating

• Largest token accounts are accounts, not beneficial owners. Pools, bridges, and custodial accounts sit at the top of most lists.
• DexScreener refuses datacenter traffic. Run the server from a laptop and it answers; run it from a cloud host and pool data falls back to the price service, which reports liquidity, 24h volume, and a holder count but no pool breakdown and no transaction counts. Tools say when a number came from the standby.
• A republished reference price can be stale outside market hours, and that alone can look like a premium.
• Aggregated venue data covers listed pools. Anything traded over the counter or internally is invisible here.
• This is informational research produced by software. It is not financial advice, and it can be wrong.

Development

npm installnpm run typechecknpm testnpm run inspector

Tests cover the analysis rules, the control translations, the registry, and the history store. They run against the build, so npm test builds first.

Releasing

The package publishes to npm as @metisagent/mcp, and server.json describes it for the MCP registry. The registry proves ownership by matching mcpName in package.json against the server name in server.json, so those two strings have to stay in step.

1. Bump the version in package.json and in both places in server.json, then add a CHANGELOG.md entry.
2. npm run verify-assets to confirm the registry still matches the chain.
3. npm test.
4. npm publish --access public.
5. mcp-publisher login github then mcp-publisher publish to list the release in the MCP registry.

License

MIT