GBLIN MCP Server

Market risk regime and treasury tools for AI agents on Base mainnet. A Model Context Protocol (MCP) server that answers two questions an autonomous agent has to ask before it moves money:

1. Is the market risk-off right now? Read the live regime — calm | elevated | crash — derived from an on-chain Crash Shield. Free, no API key, no wallet. If you need to prove to a counterparty that you checked, buy a signed EIP-712 risk attestation as portable proof-of-diligence.
2. Where does idle USDC sit between jobs? In GBLIN, a collateral-backed cbBTC/WETH/USDC index that mints and redeems at NAV, with a Just-In-Time swap back to USDC the millisecond the agent needs to pay an x402 invoice. No lockup, unsigned calldata only — your wallet signs, we never hold keys.

📖 Full documentation and Quick Start: gblin.digital/agents

⚡ Copy-paste starter examples: examples/ — ElizaOS · AgentKit / TypeScript · Claude / any MCP client — each with the recommended treasury-policy system prompt and safe-default security env.

▶️ Runnable full-cycle demo (read-only, no keys): npx tsx examples/full-cycle.ts — live against Base: treasury state → risk regime → attestation verified offline → invest calldata → JIT redemption calldata. Source.

🌐 Hosted MCP (Streamable HTTP, no install): https://gblin-mcp.gblin-mcp-worker.workers.dev/mcp — 8 free tools in a dot-notation tree, no auth, no session: risk.regime, risk.attestation_sample, protocol.stats, protocol.info, coherence.report, plus AI Action Receipts (receipts.seal demo 5/day/IP, receipts.get, receipts.verify — pure-math verification, no trust in this server). Two prompts: risk_gate, seal_and_verify. The pre-rename flat names still work as unlisted aliases until 2026-11-21. Docs are MCP resources, not tools: gblin://howto/attestation, gblin://howto/seal, gblin://limits, gblin://keys (verifier keys + key-rotation policy). Nothing is paid over MCP.

GET-only audit (for reviewers and clients that cannot POST): /meta (counts + manifest_hash) · /tools.json · /resources.json · /conformance · /v1/verify/:index (per-check booleans + on-chain anchor consistency). This hosted surface is deliberately different from the npm stdio package below (13 tools, treasury/governance included). Also on Smithery. Source in worker/.

ElizaOS Plugin

For agents running on ElizaOS, install the companion plugin — now listed in the official ElizaOS plugin registry (v0.4.0, security-hardened):

npm install plugin-gblin
# or: elizaos plugins add gblin

It exposes four native Actions (CHECK_GBLIN_TREASURY_HEALTH, INVEST_IDLE_USDC_GBLIN, RESCUE_USDC_FROM_GBLIN, GET_GBLIN_RISK_ATTESTATION) and a Provider that injects live NAV + Crash Shield status into the agent context on every loop.

→ Full Eliza plugin docs

AI Action Receipts — a witnessed transparency log for what your agent did

The problem this attacks is the #1 barrier to AI adoption in 2026: nobody can prove, after the fact, exactly what an AI did (surveys: workers burn 2–4 h/week verifying AI output; 70% of orgs say they cannot govern their agents). Our answer is the smallest honest primitive: a public, append-only RFC 6962 transparency log for AI actions. Input and output go in as hashes only (never content); the short action label, agent_id, tool and meta strings you send are published in the public log — put identifiers there, never secrets. You get back a portable receipt any third party can verify offline, forever.

receipt = canonical payload
        + Ed25519 signature            (key: gblin.digital/receipts-log)
        + RFC 6962 inclusion proof     (leaf → Merkle root)
        + C2SP signed checkpoint       (origin, tree size, root)

Canonicalization is frozen as gblin-canonical-json/1: object keys sorted by UTF-16 code unit, no whitespace, JSON.stringify semantics for primitives, recursion for objects/arrays. Test vector: payload {"b":1,"a":null} → canonical {"a":null,"b":1} → leaf = SHA256(0x00 || canonical_bytes). The receipt signature is Ed25519 over "gblin-receipt/v1\n" + canonical.

• Seal (paid, unlimited): POST https://gblin.digital/api/x402/seal — $0.01 USDC via x402.
• Seal (demo, 5/day/IP): POST <worker>/v1/seal-demo or hosted MCP tool receipts.seal (verify with receipts.verify).
• Read free forever: <worker>/v1/receipt/:index · /log · /log/checkpoint · /log/proof/:index · human page /receipt/:index.
• Daily EAS anchor on Base of the tree root (verifiable on base.easscan.org, schema 0x9f433a96…, promiseId keccak256("gblin-receipts-log")).
• Offline verifier, zero dependencies: verify-receipt.mjs — node verify-receipt.mjs receipt.json.

A seal proves existence and time in a signed append-only log whose root is anchored daily on Base. The checkpoint is currently signed by the log operator (us); independent witness cosigning is an open invitation — we already cosign a third-party transparency log (C2SP tlog-witness) and will list any witness that cosigns ours. Until then, honest wording: operator-signed, chain-anchored. It is not a compliance certificate and not an endorsement of the content. Worker: <worker> = https://gblin-mcp.gblin-mcp-worker.workers.dev.

Coherence Proof — verify GBLIN keeps its promises

GBLIN pre-registers hash-pinned public promises, then runs an automaton that probes them every 10 minutes and seals each closed day as an EAS attestation on Base. Reading is free forever; the paid service is being observed — the certifier submits itself to its own instrument first.

• Free report: /coherence (JSON) or the hosted MCP tool coherence.report.
• Live promises: P1 (uptime of the paid Risk Attestation endpoint) and P2 (honesty of the public agent-economy counters, with our own wallets disclosed and excluded).
• On-chain: each closed day is sealed as an EAS attestation on Base by a dedicated observer wallet. GBLIN is a discoverable ERC-8004 agent (#59286).

AI assistant discovery — gblin-init

Drop an AGENTS.md treasury policy into your repo so every modern AI coding assistant (Claude Code, Cursor, Windsurf, GitHub Copilot, Codex) automatically discovers GBLIN when it works on your codebase:

npx -p @gblin-protocol/mcp-server gblin-init

The command is idempotent and safe:

• If AGENTS.md does not exist → creates it from the canonical template at gblin.digital/AGENTS.template.md.
• If AGENTS.md exists → appends a clearly delimited GBLIN block at the bottom; your own content is never touched.
• Re-running is a no-op (detected via stable HTML markers).
• --dry-run prints what it would do without writing.
• --force refreshes the block to the latest version.

It also runs offline thanks to a bundled fallback template.

Postinstall hint

When the package is installed as a regular dependency (npm i @gblin-protocol/mcp-server), a short tip is printed once suggesting the command above. No files are written during postinstall — security-friendly and Socket-compliant. Set GBLIN_SKIP_HINT=1 to silence it.

Trust & Governance

GBLIN V6 is governed by a 48h Timelock Controller — every admin operation (parameter change, oracle update, ownership transfer) is enforced on-chain to wait 172,800 seconds before execution. Agents and integrators can verify this directly on BaseScan.