@mpilot/agentkit

Coinbase AgentKit adapter for the framework-agnostic @mpilot/tools registry, built on AgentKit's customActionProvider escape hatch — not the documented @CreateAction decorator path. No class hierarchy, no decorator syntax, no reflect-metadata import, no experimentalDecorators in your tsconfig.

import { AgentKit } from '@coinbase/agentkit';
import { getConciergeActionProvider } from '@mpilot/agentkit';

// Any `(agent) => ConciergeTool[]` factory works — the @concierge provider
// packages each export one (aave, dex, susde, usdy, meth, lifi, erc8004):
const provider = getConciergeActionProvider({ chainId: 5000 }, [aaveTools, dexTools]);

const agentKit = await AgentKit.from({
  walletProvider,
  actionProviders: [provider],
});

Omitting the factories yields a provider with zero actions; registry validation errors (duplicate tool names, schema violations) propagate unchanged. Action results are stringified with bigintSafeStringify (AgentKit's Action.invoke contract is Promise<string>, and provider tools return wei-scale bigints that bare JSON.stringify throws on). Tool rejections pass through by identity; serialization failures rethrow with the tool's name attached and the original error as cause.

AgentKit behaviors you should know (verified against 0.10.4)

• One provider per process — enforced. AgentKit registers custom-action metadata on the shared CustomActionProvider class and resolves getActions() from it at call time. A second registration silently rebinds dispatch for overlapping tool names (last wins) and silently merges disjoint actions into every provider's getActions() (union leakage) — both dispatch actions against the wrong agent context, so getConciergeActionProvider throws if custom actions already exist in the process. The guard fails closed — an unrecognized metadata shape (an AgentKit bump changing the registration model) throws rather than silently disarming — and zero-action providers stamp a sentinel so they count as the one provider. Raw customActionProvider calls elsewhere bypass this guard.
• Action names are prefixed. AgentKit stamps ${ClassName}_${name} on every custom action, so proposeAction surfaces to the model as CustomActionProvider_proposeAction.
• Telemetry can crash your process. Every action invocation fires AgentKit's own un-awaited fetch to cca-lite.coinbase.com, with no rejection handler and no off-switch upstream. On Node >= 22 (this package's engine floor) an unreachable endpoint — air-gapped or egress-filtered deployments — surfaces as an unhandled rejection that terminates the process on every action invocation, after the tool itself already ran. Allow that egress, intercept fetch, or install a targeted unhandledRejection handler; stub fetch in tests.
• Actions run in the agent's chain context, not the wallet's. Chain gating (supportsNetwork) runs once at registration against agent.chainId; the AgentKit walletProvider's network is never consulted (our closures are deliberately unary), and upstream hardcodes custom providers' supportsNetwork to true for every network.
• graphql peer gap. @coinbase/agentkit's barrel eagerly imports graphql-request, which peers on graphql without AgentKit declaring it. Under pnpm with auto-install-peers=false you must install graphql yourself or the import throws Cannot find module 'graphql'.

Validation semantics

AgentKit's wrapper calls the tool's own zod inputSchema.parse(args) before delegating, so ConciergeTool.invoke receives the parsed value (defaults applied, unknown keys stripped) — the same invariant as every other Concierge adapter. AgentKit core only ever touches .parse, which is why zod-4 schemas work inside zod-3-era AgentKit; the cast lives in getConciergeActionProvider, scoped to the schema field, and is pinned by tests. Scope caveat: bridge packages that convert action.schema to JSON Schema with zod-3 tooling (e.g. AgentKit's framework extensions) are NOT verified with zod-4 schemas and may silently advertise an empty parameters schema — wire the provider into AgentKit.from(...) and drive the model loop yourself. outputSchema is not enforced on return values — AgentKit's Action contract has no return-shape slot; output validation belongs to the tool.

toAgentKitAction(tool) is exported for composing individual tools into your own customActionProvider array (note that this path bypasses the one-provider-per-process guard, and the returned invoke is pre-validation — only call it through AgentKit). It throws a TypeError on .transform()/.pipe() input schemas (downstream action→model bridges convert the schema to JSON Schema, where a pipe may advertise its output segment) and on non-object schemas.