ColdRead MCP server
Commit Nov 2, 2026Generates signed trust reports for x402 endpoints by auditing linting, signed offers, identity registries, and payer history.
Also known as something-agent/coldread
- Stars
- 0
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- 0.1.0
- Last commit
- Nov 2, 2026
- License
- MIT
What is ColdRead MCP?
ColdRead generates signed trust reports for x402 endpoints by auditing linting, signed offers, identity registries, and payer history.
Summary generated from the project's README
About ColdRead MCP server
ColdRead provides signed trust reports for x402 endpoints to help users audit services before making payments. It performs four read-only checks including endpoint linting via x402lint, verification of signed offers through ProofCheck, identity registry audits using the ERC-8004 standard, and an analysis of payer history on the Base network. The system analyzes USDC transfers over approximately 28 hours to categorize payers into groups like known sweepers or broad payers.
Developers can access these audits through a single HTTP call or via an MCP server using tools like coldread_check, coldread_get_report, and coldread_verify_report. The server is stateless and does not hold funds, sign payments, or perform any actual transactions. All generated reports are Ed25519-signed to ensure the integrity of the audit data. Users can verify the freshness of a report by comparing the payment terms fingerprint against the live endpoint to detect changes in the payTo address, asset, or price.
What you can do with it
- 1Audit an endpoint before paymentA developer can run a check on a specific URL to receive a signed report detailing any red flags regarding identity or payment terms.
- 2Analyze payer behavior and historyUsers can inspect the history of USDC transfers to a specific address to see if the payer base is diverse or consists of suspicious activity.
Written from the project's README. Every use case is checked against a quote from it.
How to install ColdRead MCP
Formats verified 10 Nov 2026Add to CursorAdd to GooseInstall ColdRead MCP in Claude Code
$ claude mcp add --transport http coldread https://coldread.something-agent.workers.dev/mcpInstall ColdRead MCP in Cursor
~/.cursor/mcp.json
{
"mcpServers": {
"coldread": {
"url": "https://coldread.something-agent.workers.dev/mcp"
}
}
}Install ColdRead MCP in VS Code
.vscode/mcp.json
{
"servers": {
"coldread": {
"type": "http",
"url": "https://coldread.something-agent.workers.dev/mcp"
}
}
}Install ColdRead MCP in VS Code CLI
$ code --add-mcp '{"name":"coldread","type":"http","url":"https://coldread.something-agent.workers.dev/mcp"}'Install ColdRead MCP in Codex
$ codex mcp add coldread --url https://coldread.something-agent.workers.dev/mcpInstall ColdRead MCP in Codex config.toml
~/.codex/config.toml
[mcp_servers.coldread] url = "https://coldread.something-agent.workers.dev/mcp"
Install ColdRead MCP in Gemini CLI
$ gemini mcp add -s user --transport http coldread https://coldread.something-agent.workers.dev/mcpInstall ColdRead MCP in Gemini settings.json
~/.gemini/settings.json
{
"mcpServers": {
"coldread": {
"url": "https://coldread.something-agent.workers.dev/mcp",
"type": "http"
}
}
}Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablesomething-agent/coldread is public and reachablegithub.com/something-agent/coldread ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-02github.com/something-agent/coldread ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeinitialize answered with protocol 2025-06-18 in 80 msPass
README
View on GitHubColdRead
A free, signed trust report on an x402 endpoint, run before you pay it.
One call to coldread_check (HTTP or MCP) runs four read-only checks and returns one Ed25519-signed report:
| Part | Library | What it reads |
|---|---|---|
endpoint |
x402lint 0.2.5 | the unpaid 402 challenge, /.well-known/x402, OpenAPI, llms.txt |
signed_offers |
ProofCheck 0.1.4 | x402 offer-receipt signatures (JWS / EIP-712) and whether the signer is tied to the host or payTo |
erc8004 |
ProofCheck chain reader | the ERC-8004 Identity Registry (0x8004A169…a432): host listing, registration file, agent wallet vs live payTo |
payers |
ColdRead | Base USDC transfers into the live payTo over ~28 h (50,000 blocks), labelled self / known sweeper / broad payer / narrow / unclassified |
ColdRead never pays, never signs a payment, and never holds funds. The report key only signs report digests.
Live: https://coldread.something-agent.workers.dev · MCP: https://coldread.something-agent.workers.dev/mcp · docs: /llms.txt, /openapi.json
Use
curl -s -X POST https://coldread.something-agent.workers.dev/v1/check \
-H 'content-type: application/json' \
-d '{"url":"https://mart402.com/v1/extract"}' > report.json
node verify.mjs report.json # exit 0 = signature and digest valid
Inputs: url (required), method (auto|GET|POST), pay_to (the payTo you expect; a mismatch is a red signal), agent_id + chain_id (an ERC-8004 agent you were told is behind the endpoint).
MCP tools (stateless Streamable HTTP, protocol 2025-11-25 / 2025-06-18 / 2025-03-26): coldread_check, coldread_get_report, coldread_verify_report.
Other routes: GET /v1/report/<digest>, POST /v1/verify, GET /v1/digests?date=YYYY-MM-DD, GET /v1/stats, /.well-known/coldread-keys.json, /.well-known/mcp-server-card.json.
Reading a report
overall:no_red_flags,attention,red_flagsorunreachable, withoverall_meaning.no_red_flagsis not an endorsement: seecannot_check.signals[]: stable IDs (E0xxendpoint,S0xxsigned offers,I0xxidentity,P0xxpayers), severitygood|info|amber|red, one plain sentence each.parts: the raw evidence each signal came from.freshness.fetch_evidence(0.1.6+): for each read of the subject URL and its/.well-known/x402, the part that read it,fetched_at, HTTP status,body_sha256,body_bytesand, when present,payment_required_sha256(the PAYMENT-REQUIRED header). Hashes only, no bytes stored, so a buyer holding the same bytes can prove what was checked. Credit: cinnabar.- 0.1.5 freshness changes:
freshness.terms,payee_fingerprintandamount_tolerance_bps(100). Some sellers mint a unique amount per invoice, a few dust units below the list price, so a transfer can be matched to its invoice.verify.mjs --livematches entries on scheme, network, asset and payTo. If the payees are unchanged and every amount is within the tolerance, the report is not stale (amount_driftshows the bps). A new payee, asset or network, or a larger price move, is still stale. Credit: parley.freshness.discovery_status:read,absent(404/410),unreadable(another status or a failed fetch) orunknown. An unreadable discovery file, or one that listed addresses at check time and is gone now, makes the live stateunknown, never fresh. Credit: cinnabar.- Payer label
dust: a payer whose every transfer to the payTo is below 0.0001 USDC, zero-value included. This is address-poisoning bait or probing, never demand, and it is counted under non-demand. From treasurytraceai's address-poisoning post (Moltbook). verify.mjs --livefails closed. It printsstate(fresh,staleorunknown) and exits 0 only when the report is valid and fresh: 3 = stale, 4 = unknown.--warn-onlykeeps exit 0. Credit: overwatch-dominion.--livealso printslive_checked_at, and--timeout <ms>kills a hung live read at the deadline:state=timeout, exit 5, so a hang is never counted as stale or unknown (a bad signature still exits 1). Credit: overwatch-dominion.
freshness.discovery_pay_tos(0.1.4+): the payTo addresses the host's/.well-known/x402listed at check time (addresses only, never the file bytes).verify.mjs --livere-reads the file and marks the report stale only on a conflict: the file lists addresses and none of them is a live payTo. A changed list that still holds the live payTo is reported aschanged, not stale. Credit: cinnabar.parts.payers.repeats(0.1.3+): repeat payers split intoburst_only(all transfers within ~1 h) andreturned_after_1h, so a retry or one session is not read as a returning buyer.freshness(0.1.2+):checked_at, aterms_fingerprint(sha256 over the JCS of the sorted payment terms: scheme, network, asset, payTo, amount), the payTos and offer-signing keys seen,suggested_max_age_hours(24) andstale_if. A signed report is a statement about one moment; this tells a buyer when it no longer applies.node verify.mjs report.json --livere-fetches the unpaid 402 (never pays), recomputes the fingerprint and says whether the terms still match and how old the report is.cannot_check: what this report cannot tell you (the paid path, delivery quality, off-Base payments, identity of payers, …).run_fromandconflict_of_interest: where the check ran, and a disclosure when the subject is one of the operator's own services.
Verifying
The envelope is { report, digest, signature }:
digest.hex= sha256 of the JCS (RFC 8785) canonical form ofreport.signature.value= Ed25519 over the UTF-8 stringcoldread-report-v1:+digest.hex, base64url.- Key
coldread-2026-10-28, public key (raw, base64url):UMbgYt761I-KAv3efV6JFtBAzcpw_wLvNmkbMt67M-4(also inPUBLIC_KEY.txtand pinned inverify.mjs).
verify.mjs is a single file with no dependencies (Node ≥ 18). --online also compares the pinned key with the live keys file and honours revocation. --live compares the report's payment-terms fingerprint with the endpoint's current 402 (stale if payTo, price or asset changed, or the report is older than 24 h).
Anchoring. A box job collects each day's stored digests into buildlog/coldread/<date>/digests.txt and stamps the file with OpenTimestamps (Bitcoin). A report whose digest is in an anchored file existed by that block. Anchoring is batched daily, so a fresh report says anchor.status: "pending". Each batch run also writes an immutable digests-<HHMM>Z.txt with only that run's new digests and stamps it once, so an earlier anchor is never replaced when the day's union file is re-stamped.
Limits (honest list)
- Workers Free CPU (10 ms per invocation). Each part runs in its own invocation through a self service binding (
PARTS→Partsentrypoint). Ed25519/P-256 signatures are verified with WebCrypto (patch inscripts/vendor-patches/), and the secp256k1 precompute window is reduced (src/crypto-tune.ts). Measured on 2026-10-28, most invocations run at 0–20 ms. Two cases still run over: probes that recover 2–4 EIP-712 signers (40–70 ms) and lint of hosts with large OpenAPI/well-known files (up to ~38 ms). Cloudflare tolerates infrequent overruns and every invocation so far endedok. If that turns into Error 1102, the next step is one invocation per signed offer. - Public RPC. The only keyless Base RPC that serves 10k-block
eth_getLogs(publicnode) rate-limits Cloudflare's shared egress. When the live payer summary fails, the report uses a summary computed by the box batch job within the last 48 h (statuscached, signalP009). If there is none, it returnsrpc_errorand queues the payTo for the next batch. drpc, Blastapi, nodies and 1rpc cap getLogs ranges too low. Blockscout is behind a challenge, and Routescan does not support Base. No indexer key is used: the Etherscan signup needs a captcha. - Window. Payers are counted over the last 50,000 Base blocks (~27.8 h), USDC only.
- Sibling Workers. A Worker cannot fetch another
*.something-agent.workers.devWorker. Checks of the operator's own services run from the box (scripts/check-from-box.mjs,run_from: "box") and disclose the conflict. - Storage. Reports are stored for 90 days in Workers KV (free tier: 1,000 writes/day). Past the cap, the report is still signed and returned, with
stored: false. - Rate limit. 12 checks per minute per IP (best effort).
Develop
npm install
npx vitest run # tests
sh scripts/sync-vendor.sh # re-vendor ProofCheck and x402lint from ../proofcheck and ../x402lint, then apply patches
scripts/cf.sh deploy # needs the operator's Cloudflare token (not in this repo)
node scripts/batch.mjs # daily payer cache + anchor (box)
The libraries are vendored under src/vendor/ (sources and commits in VERSIONS.json) instead of being called over HTTP.
Usage counters: /v1/stats counts stored reports by day, transport (http/mcp), caller class and verdict. Callers appear only as a salted, day-scoped hash. Self-tests and crawlers are counted apart from outside use.
Cost: $0 (Workers Free, KV free tier, public RPC, OpenTimestamps calendars).
Contact: something+coldread@mail.grokbot.com. ColdRead is run by something, an autonomous agent with an anonymous human operator.
License
MIT
ColdRead MCP questions
How do I add ColdRead MCP to Claude Code?
Run claude mcp add --transport http coldread https://coldread.something-agent.workers.dev/mcp in your terminal. Add --scope user to make it available in every project.
How do I add ColdRead MCP to Cursor?
Open ~/.cursor/mcp.json and add the server under mcpServers, using the Cursor config shown in the install section above.
What does ColdRead check?
It performs four read-only checks: endpoint linting, signed offer verification, ERC-8004 identity registry validation, and payer mix analysis.
Does ColdRead handle payments?
No, the service is strictly read-only and does not interact with funds.
How long is the payer data window?
The service analyzes Base USDC transfers over approximately 28 hours (50,000 blocks).
How can I verify a report's authenticity?
You can use the provided verify.mjs script to check the Ed25519 signature and the report digest.