🐴 Gulltoppr

🐴 gulltoppr

*A real session: an unverified MEV bot that traded seconds earlier (no source, no ABI anywhere) resolved to a full interface in two MCP tool calls.*

The REST engine for gulltoppr: the resolution ladder + verb surface that lets an AI agent go from (chain, address) to a correct, simulated, safe contract interaction. This is "the engine" of the four faces (REST → MCP → SDK → Skill); see ../SPEC.md for the full contract and ../IDEATION.md for the strategy.

TypeScript + viem + Hono. The heimdall decompile rung is delegated over HTTP to gulltoppr (kept out-of-process by design).

Run

npm install

npm run dev # REST engine: tsx watch on http://localhost:8787 npm run mcp # MCP server: stdio tools for agent clients npm run mcp:http # MCP server: Streamable HTTP (remote agents) npm run typecheck # tsc --noEmit npm test # vitest unit tests (cache, chains, ladder helpers, args, errors) npm run verify # root typecheck/test/audit + SDK build/test/audit npm run docker:build # build the REST engine deployment image npm run docker:build:mcp # build the Streamable HTTP MCP deployment image npm run docker:smoke # start both deployment images and check health/discovery/registry metadata npm run test:live # opt-in live contract smoke tests (RPC/decompiler/network)

Set LIVE_ENGINE_BASE_URL=https://api.gulltoppr.dev with npm run test:live to smoke a deployed engine instead of the local in-process server. The scheduled live GitHub Action runs the local path; manual dispatch can set the same deployed URL.

Env

Endpoints (SPEC §4)

GET routes set explicit Cache-Control: verified ABI responses cache longest, proxy ABI responses are short-lived, transaction decodes are immutable, the OpenAPI contract is cacheable, and operational endpoints such as /health and /v1/metrics are no-store. Rate-limited routes expose RateLimit-Limit, RateLimit-Remaining, RateLimit-Reset, and Retry-After on 429 responses. ABI resolves also return X-Source, X-Confidence, X-Cache, X-Elapsed-Ms, and X-ABI-Included.

The registry (selector commons)

The engine seeds an open selector→signature registry as a byproduct of resolution:

• Every verified resolution (Etherscan/Sourcify) harvests ground-truth selector → signature pairs for functions, events (full 32-byte topic0, collision-free), and errors. Proof grade: verified-source.
• Resolutions are also indexed by skeleton hash (runtime bytecode with the solc metadata trailer stripped), so byte-identical clones resolve via a new bytecode-match rung without re-running the ladder. Verified claims are capped to partial for clones (this address's source was never verified), and provenance.bytecode_match points at the original chain/address/source/confidence that supplied the reused ABI.
• Decompiled ABIs get Unresolved_ names replaced from proven registry entries, and (when ANTHROPIC_API_KEY is set) a fire-and-forget propose-and-verify pass asks Claude for candidate signatures and accepts only those where keccak256(sig)[:4] reproduces the selector: proof grade keccak-proven (signature proven; semantics still inferred).
• Public 4byte fallback labels are also selector-matched locally before use, but remain unproven labels and never enter the commons as proof.

Only the engine's own pipeline writes to the registry; no open submissions (that's how 4byte got collision-poisoned).

The accumulated data is published as a CC0 dataset: evm-abi-commons (regenerate any time from GET /v1/registry/export; the response is NDJSON and includes X-License: CC0-1.0 plus a license Link header). Lookup/export entries include proof grade and, for harvested verified-source entries when known, the source chain and address. SDK users can call lookupSelector, registryStats, and exportRegistry directly.

{chain} is any alias from GET /v1/chains (backed by viem/chains) or a numeric id. Chain entries include testnet and has_default_rpc so agents can decide when to ask for rpc_url; q matches ids, names, aliases, native symbols, and multi-word searches such as bnb chain. Pass ?rpc_url= to override the RPC (required for chains with no default, e.g. local/31337; this is how any EVM chain works before it has a built-in alias).

For agent contexts, prefer include_abi=false on resolve_abi unless you need the raw JSON ABI. The compact response preserves interface, provenance, proxy, token, and abi_for, and marks abi_omitted: true. For large contracts, add method_q, method_kind=read|write|all, and method_limit to return only the manifest methods relevant to the user's intent.

curl localhost:8787/v1/ethereum/0xC02aaA39b223FE8D0A0e5C4F27eAD9083C756Cc2/abi

curl -X POST localhost:8787/v1/ethereum/0xC02aaA39b223FE8D0A0e5C4F27eAD9083C756Cc2/prepare \ -H 'content-type: application/json' \ -d '{"function":"approve","args":["0x1111111254EEB25477B68fb85Ed929f73A960582","1000000000000000000"],"from":"0xd8dA6BF26964aF9D7eEd9e03E53415D37aA96045"}'

prepare_tx.safety.signing_recommended gates the hand-off. If simulation fails, risk_level is blocked, deeplink is empty, and wallet_request is omitted. When signing is recommended, wallet_request is an EIP-1193-shaped eth_sendTransaction payload with hex JSON-RPC quantities for wallet/explorer/app integrations. Decompiled or selector-only writes are high risk and require explicit user confirmation of selector + intent. Positive token/NFT spender approvals are medium risk with spending_approval. Token/NFT outflows from the transfer source, whether trace-derived or inferred from a standard token/NFT transfer, are medium risk with asset_outflow. Clients should show both before hand-off.

Integration recipes for wallets, block explorers, coding agents, and MCP clients are published at docs/integrations.md.

MCP server (SPEC §5)

npm run mcp starts a stdio MCP server exposing the seven core verbs plus read-only utility tools for chains, selector commons, and runtime metrics. The tools are a thin adapter over the deployed REST engine (ENGINE_URL), so the MCP shares the engine's persistent cache and Etherscan key; no duplicated resolution or secrets. Tool descriptions bake in the non-custodial hand-off model (prepare_tx never signs), lead with provenance warnings for partial/proxy/bytecode-match/decompiled ABI results, and require prepare_tx.safety.signing_recommended before a signing deeplink or wallet request is handed to the user. JSON MCP tools expose output schemas and structuredContent; clients can branch on provenance, decoded calls, simulations, selector results, metrics, and safety without scraping text.

Wire it into an MCP client (Claude Desktop / Claude Code mcp config):

{

"mcpServers": { "gulltoppr": { "command": "npm", "args": ["run", "--silent", "mcp"], "cwd": "/home/ubuntu/repos/abi-agent", "env": { "ETHERSCAN_API_KEY": "" } } } }

Core tools: resolve_abi, read_contract, encode_call, simulate, prepare_tx, decode_tx, resolve_name. Utility tools: list_chains, lookup_selector, registry_stats, export_registry, runtime_metrics. All are read-only-annotated except prepare_tx (non-destructive: returns an unsigned hand-off, signs nothing).

Remote (Streamable HTTP)

For agents that can't run a local stdio server, the same MCP is hosted over HTTP at https://mcp.gulltoppr.dev/mcp (npm run mcp:http locally; stateless). Point an HTTP-capable MCP client at that URL:

{ "mcpServers": { "gulltoppr": { "url": "https://mcp.gulltoppr.dev/mcp" } } }

Tool registration is shared (src/mcp-server.ts) between the stdio entry (mcp.ts) and the HTTP entry (mcp-http.ts), deployed via Dockerfile.mcp / fly.mcp.toml. MCP directory metadata lives in server.json and advertises the same remote URL, repository, homepage, and icon. The remote MCP service also serves the same metadata at https://mcp.gulltoppr.dev/server.json and https://mcp.gulltoppr.dev/.well-known/mcp-server.json.

npm SDK

A typed client over this REST surface lives in sdk/ (gulltoppr): new Gulltoppr({ baseUrl }).resolveAbi(...) / .read(...) / .prepareTx(...), plus a contract() helper. It's the third face (after REST and MCP) and the basis for wallet, explorer, and app integrations. See sdk/README.md.

Depl