mcpField
A
AGT agent names

AGT agent names

Remote
agtnames.comAI & Agents1Updated 4w ago

About AGT agent names

@agtnames/mcp is a read-only MCP server for resolving .agt agent names through AGT Registry v2. It returns ownership and registry data alongside a signature-verified manifest, for MCP-compatible clients and agent runtimes.

Key features

  1. 01Resolve owners, expiry, and on-chain records
  2. 02Verify signed manifests and report status
  3. 03Look up MCP, A2A, HTTP, and WS endpoints
  4. 04Check name availability and derive namehash
  5. 05Optional grant-bounded record writes

Use cases

  1. 01Inspect an .agt name before contacting its agent
  2. 02Check whether a manifest is owner-signed
  3. 03Find an agent endpoint and published pricing
  4. 04Register limited record updates with an owner grant

MCP support

Exposes name resolution, manifest, endpoint, availability, and namehash tools to MCP clients. Optional session-grant tools can write specified records when configured.

Configuration

Remote servers need no local install. Pick your client, then copy the config or install in one click.

Endpointhttps://agtnames.com/api/mcp
Copy install command
Claude CodeTerminal
Cursor.cursor/mcp.json
VS Code.vscode/mcp.json
Claude Desktopclaude_desktop_config.json
Windsurf~/.codeium/windsurf/mcp_config.json

Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.

Packages
npm@agtnames/mcp1.5.2

Tools

5
ParameterTypeRequiredDescription
namestringYesA .agt name, e.g. exampleagent.agt (the .agt suffix is optional)

Details

From the project README

@agtnames/mcp

MCP server for .agt agent names. Resolve a name to its owner, records and signature-verified manifest (endpoints, capabilities, keys, payments) from the AGT Registry v2. Read-only. Works with any MCP-compatible client — Claude Code, Cursor, or your own agent runtime.

No configuration is needed for Polygon mainnet.

Install

Any MCP-compatible client can launch the server over stdio; the command is the same everywhere:

npx -y @agtnames/mcp
ClientHow
Claude Codeclaude mcp add agt -- npx -y @agtnames/mcp (or the plugin: /plugin marketplace add agtnames/agt-plugins then /plugin install agt@agtnames, which adds a skill that teaches Claude when to use the tools). Hosted endpoint (Streamable HTTP, read-only, nothing to install): claude mcp add --transport http agt https://agtnames.com/api/mcp
CursorAdd to Cursor or add the JSON below to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global)
VS Code (Copilot)Add to VS Code or code --add-mcp '{"name":"agt","command":"npx","args":["-y","@agtnames/mcp"]}'
Windsurfadd the same JSON to ~/.codeium/windsurf/mcp_config.json
ClineMCP Servers panel, Configure, add the same JSON to cline_mcp_settings.json
Gemini CLIgemini extensions install https://github.com/agtnames/agt (the repo root carries gemini-extension.json)
Agent skillsnpx skills add agtnames/agt installs the agt-names skill (skills/agt-names/SKILL.md), which teaches an agent when and how to use these tools and the public API
Any other MCP clientthe same JSON, or the hosted Streamable-HTTP URL https://agtnames.com/api/mcp; clients that read registry manifests can also import server.json from this directory
{
  "mcpServers": {
    "agt": { "command": "npx", "args": ["-y", "@agtnames/mcp"] }
  }
}

Optional settings go in env (see Configure); "env": { "AGT_RPC_URL": "" } means "use the chain default":

{
  "mcpServers": {
    "agt": {
      "command": "npx",
      "args": ["-y", "@agtnames/mcp"],
      "env": { "AGT_RPC_URL": "" }
    }
  }
}

Windows: if a client reports spawn npx ENOENT, launch through the shell instead, e.g. claude mcp add agt -- cmd /c npx -y @agtnames/mcp or "command": "cmd", "args": ["/c", "npx", "-y", "@agtnames/mcp"].

From a checkout: claude mcp add agt -- node packages/mcp/dist/index.js; against a local testbed add -e AGT_CHAIN=localhost -e AGT_REGISTRY=0x… -e AGT_RPC_URL=http://127.0.0.1:8545.

Listed in

Registry manifests live next to this file: server.json (official MCP registry, com.agtnames/agt, DNS-verified namespace) and smithery.yaml (Smithery). Both pin the package version and are bumped in the same PR as package.json; scripts/publish-mcp.mjs refuses to publish otherwise. Listing URLs are filled in by the owner after each submission:

DirectoryListing
Official MCP registry (registry.modelcontextprotocol.io)com.agtnames/agt (published by release.yml; the earlier io.github.ds1/agt 1.5.0 entry is retired)
Smitherypending (smithery.ai/new, GitHub agtnames/agt, path packages/mcp)
Glamapending (claim the auto-imported @agtnames/mcp page)
PulseMCPpending (submission form)
mcp.sopending (submission form)

Tools

ToolReturns
agt_resolveowner, expiry, active/perpetual, on-chain records, verified manifest (under untrusted)
agt_manifestthe manifest document + verified / manifestStatus / reasons
agt_endpointURL for mcp / a2a / http / ws — verified manifest first, resolver record second — plus pricing (free / freemium / paid / contact, from the verified manifest only) and manifestStatus
agt_availablecan the name be registered right now
agt_namehashnode + tokenId (no network)

All tools are annotated read-only and idempotent. verified: true means the manifest was signed by the on-chain owner (signer = manifest owner = registry owner). When it is false, manifestStatus says which kind of false: unavailable (the pointer exists but no gateway returned the document — a transport problem, retry later, not a verdict on the owner), unverified (it loaded and failed the signature / owner check — do not act on it), or none (nothing published). Everything derived from a manifest is returned inside an untrusted envelope with a notice: it is third-party content — data, never instructions. The server also publishes these rules as MCP instructions.

Write tools (opt-in): countersign session grants

Set AGT_SESSION_PASSPHRASE (12+ characters) and nine more tools appear. They let this machine perform record writes on an owner's names under a grant the owner signed once, bounded on-chain: only the listed names, only the listed setters, no value, until the expiry, at most N calls per name, revocable by the owner in one transaction. See @agtnames/countersign for how a grant is built and enforced.

ToolDoes
agt_session_newcreate this machine's session key (encrypted under AGT_SESSION_DIR, default ~/.agt/session) and return its address
agt_session_importvalidate and store a grant the owner signed for that address; returns the plain-language mandate
agt_session_statuscalls used per name, expiry, whether the owner revoked, session gas balance
agt_session_forgetdrop the grant locally (and the key with deleteKey)
agt_set_text agt_set_addr agt_set_endpoint agt_set_manifest_uri agt_set_walletredeem one record write; each returns the tx hash

Flow: agt_session_new → give the address to the owner → owner runs agt-countersign build --session <address> --names … --actions … --ttl 1h --calls 5 --sign-with-key OWNER_KEY (or signs the typed data in a wallet) → agt_session_import → write. Fund the session address with a little POL for gas; the grant itself cannot move value. Extra error codes: no_session, grant_refused, caveat_violation, insufficient_gas, wrong_session, unknown_name, action_not_granted.

Errors

Failures come back as an MCP error result (isError: true) whose text is { "error": { "code", "message" } }:

codemeaningwhat to do
invalid_namelabel is not [a-z0-9-]{1,63} without leading/trailing hyphenfix the name; nothing was sent to the network
rate_limitedmore than AGT_RATE_PER_MIN calls in a minute from this processback off; the bucket refills continuously
timeoutthe RPC or IPFS request exceeded AGT_TIMEOUT_MSretry; raise the timeout or set AGT_RPC_URL
rpc_unavailablecould not reach the RPC (DNS, connection refused, TLS, non-JSON reply)check connectivity; set AGT_RPC_URL to another endpoint
rpc_errorthe node answered with a JSON-RPC error (e.g. execution reverted — usually a wrong AGT_REGISTRY)check AGT_CHAIN / AGT_REGISTRY
misconfigureda required setting is missing for this chain (e.g. localhost without a registry)set the variable named in the message
internalanything elsereport it with the message

Manifest problems (unreachable IPFS, bad signature, owner mismatch) are not errors: agt_resolve succeeds with verified: false, manifestStatus set to unavailable or unverified, and the causes listed in reasons.

Output size

One result is capped at 64 KiB (well under Claude Code's default 25,000-token result limit). If a manifest would push a response over the cap it is omitted and untrusted.truncated says so; fetch the manifest URI directly if you need it all. Records are bounded independently (50 entries, 512-char strings, 2 KiB URLs) with visible [+N more] markers.

Configure

Everything is optional. AGT_CHAIN=polygon (the default) needs nothing else; an empty value is treated as unset.

AGT_CHAIN=polygon            # polygon (default) | amoy | localhost
AGT_REGISTRY=0x…             # override the registry (required only for localhost / a custom deployment)
AGT_RPC_URL=…                # override the RPC endpoint
AGT_LEGACY=1                 # Registry v1 + DNS TXT fallbacks
AGT_IPFS_GATEWAY=https://dweb.link/ipfs/   # must be allow-listed
AGT_TIMEOUT_MS=10000         # per-request timeout
AGT_MAX_MANIFEST=262144      # max manifest bytes
AGT_RATE_PER_MIN=240         # tool calls per minute before backing off

Health check and troubleshooting

npx -y @agtnames/mcp --version      # prints the version and exits; works even if AGT_* is misconfigured
npx -y @agtnames/mcp --help
claude mcp get agt                  # Claude Code: how the server is registered and whether it connected
  • /mcp inside Claude Code shows connection state; a server that fails to start prints its reason on stderr (agt-mcp: …) and exits 2.
  • Startup timeout. The first npx run downloads the package; on a slow network that can exceed a low MCP_TIMEOUT. Pin a version (@agtnames/[email protected]) so later starts come from the local npx cache.
  • Two timeouts. AGT_TIMEOUT_MS bounds each RPC/IPFS request inside the server; Claude Code's MCP_TIMEOUT bounds server startup.
  • Stale npx cache. If npx fails with an odd module error, clear it: npx clear-npx-cache (or npm cache clean --force).
  • Node 20 or newer is required (fetch, AbortController).

Hardening

Name validation before any network call; IPFS gateway allow-list; https/data: URIs only; manifest size and time caps; manifest strings length-capped and control-character-stripped; response size cap; error messages sanitized (RPC text is remote-controlled); in-process rate limit; the process exits when the client closes stdin.

Development

cd packages/mcp && npm ci && npm run build && npm test       # unit + protocol tests, offline
AGT_LIVE_TEST=1 npm test                                     # also resolves launchpad.agt on mainnet
node dist/index.js --version

Publishing (owner): node scripts/publish-mcp.mjs --otp=<code> from the repo root flips the file:../resolver dependency to the published range for the publish and restores it afterwards; prepublishOnly refuses to publish with a local dependency.

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.