nodecode-basekit

Basekit: read-only Base (chain 8453) chain reads - network and gas in dollars, wallet portfolios, transaction receipts, token facts, transfers and approvals - as six tools the model calls. A Nodecode rewrite of D0xedDevi0/hermes-basekit by D0xedDevi0.

Get it

Tell the agent "add the basekit cell", or run nodecode add nc://basekit.

Set it up

Nothing is required. Out of the box the cell reads Base through a keyless public endpoint (https://base-rpc.publicnode.com), prices with DefiLlama and falls back to Coinbase's spot ETH price. The basekit section is optional:

Member	What it is
`rpc_url`	your own node or provider (Alchemy, QuickNode, Coinbase); the public default caps `eth_getLogs` ranges
`watch_address`	the wallet `/basekit wallet` reads when no address is typed
`timeout_seconds`	one request's ceiling, 20 by default
`log_window`	the default block window for transfers and approvals, 20000 by default (at most 50000)
`basescan_api_key`	only adds `base_token`'s source-code check; without it that field says `not_configured`

/basekit with no argument shows the chain state it is reading. (help :basekit) is the model's guide to the tools.

Use it

Ask for what you want; the model picks the tool:

Tool	Use
`base_network`	head block, gas and base fee, ETH/USD, and the dollar cost of a transfer, an ERC-20 transfer and a swap
`base_portfolio`	ETH and ERC-20 balances with USD values for one address
`base_tx`	one transaction decoded: status, block and time, from/to, fee in ETH and USD, every token movement and approval
`base_token`	name, symbol, decimals, supply, owner(), real proxy detection (EIP-1967, EIP-1167, `implementation()`) and selector hints
`base_activity`	transfers in and out of one address for one token, inside a stated block window
`base_allowances`	approvals still live on one token, each re-read on-chain, with the USD at risk

From the keyboard: /basekit status, /basekit full (adds the endpoint), /basekit wallet <address>, /basekit token <address>, /basekit activity <address> <token>.

Nothing here signs. There is no private key, no transaction path and no key needed: basekit reads, and the user decides. Every answer names the block it was read at, and the two log-based tools state the block window they actually scanned - a transfer outside it is not in the answer.

It is the plugin's six tools

The original registered six agent tools, a /basekit command, a CLI command, a skill and a prompt hook. The six tools keep their names and their arguments, and the answers keep the plugin's honesty rule: the block is quoted, an unpriced token says so rather than guessing, and bytecode hints are labelled as hints.

Not ported

The pre_llm_call hook. It injected one cached line of chain state, gated on Base-related words in the user's message. A cell's round hook does not see that message here, so the same guidance lives in (help :basekit) and in the tools' own descriptions ("call base_network before quoting a fee"). A model that follows them pays the same one call and nothing on unrelated turns.
The ledger, the CLI command and the desktop pane/status chip. The ledger was a JSONL file of past reads under the plugin's data folder; the CLI and the pane were Hermes surfaces. /basekit and the tools answer the same questions without a file to keep.
The automatic window narrowing. When a busy token blew the endpoint's result cap, the original halved the log window and retried. Here a rejected window is one error line telling the model to narrow blocks or set its own rpc_url, which is one call instead of up to seven.
Multicall3. The original batched contract reads through the canonical Multicall3 deployment; this cell sends them as one JSON-RPC batch, which is the same one round trip without an ABI encoder for aggregate3.

Test

From the cells repository root:

.github/test-cell.sh tools/nodecode-basekit

Every request is a stubbed nlk:http: the RPC endpoints answer a small fixed chain and DefiLlama answers two prices, so nothing reaches the network and no key is needed.

MIT licensed, as the original is.