Coinbase Agentic Wallet - Setup Guide¶

Source: coinbase/agentic-wallet-skills via skills.sh - 27.3K combined installs across 11 indexed listings; first documented November 3, 2026 (evening sweep) GitHub: 127 stars, license MIT, last pushed 2026-08-04, 1 SKILL.md directories Category: Agent payments: sign in, read balances, send USDC and trade tokens through the Coinbase agentic wallet CLI Quality Tier: 🟡 Yellow - MIT and a small, reviewable tree, but four of the five sampled listings carry a Snyk warning and query-blockchain-data fails Gen Agent Trust Hub

Coinbase publishes this repository as Agent Skills for crypto wallet operations, built around the awal CLI distributed on npm. The tree holds a single SKILL.md directory, skills/agentic-wallet, and that one skill is a router: its frontmatter declares the shell tools the agent may call, and its body maps a capability to the reference document the agent should read before acting.

The reference set covers sign-in by email OTP, balances across Base, Polygon and Solana, sends of USDC, ETH, POL and SOL, token swaps on Base and Polygon, funding through Coinbase Onramp, the x402 paid-API flow in three parts (search, pay, monetize) and onchain queries through the CDP SQL API. skills.sh indexes 11 listing names against that one directory, so the per-capability rows in the table below are listing names, not 11 separate skills to install.

Installation¶

The README documents one command with the skills CLI:

npx skills add coinbase/agentic-wallet-skills

The README states that installed skills are automatically available once installed and that the agent uses them when it detects a relevant task. No API key is provisioned at install time: wallet access is an email OTP sign-in handled at run time by the awal CLI, which the skill drives.

For a Hermes profile, the same tree can be dropped into the profile skills directory. The reference documents are resolved relative to the skill folder, so keep references/ alongside SKILL.md:

git clone --depth 1 https://github.com/coinbase/agentic-wallet-skills.git /tmp/cbx-skills
cp -R /tmp/cbx-skills/skills/agentic-wallet ~/.hermes/profiles/corpusiq/skills/

The awal version is pinned in every code block in SKILL.md and in each reference document rather than in a package manifest. The repository ships scripts/bump-awal.js to rewrite those pins, and its own instructions require Node v22 or newer to run it. Nothing in the repository mentions Hermes, and nothing in it is chain-specific beyond the networks named in the references.

What It Provides¶

Skill	Installs	Use For
trade	3,096	Swap or trade tokens on Base or Polygon; the routing table points here for `trade.md`.
x402	3,018	Umbrella listing for the three x402 documents: bazaar search, paying an endpoint and monetizing one.
send-usdc	2,951	Send USDC, ETH, POL or SOL on Base, Polygon or Solana via `send-usdc.md`.
pay-for-service	2,946	Pay an x402 endpoint with automatic USDC payment; `x402-pay.md`.
monetize-service	2,912	Build and deploy a paid API server on x402; `x402-monetize.md`.
search-for-service	2,811	Search the x402 bazaar for paid API services; `x402-search.md`.
agentic-wallet	2,739	The installed skill itself: the router that sends the agent to the reference file matching the task.
authenticate-wallet	2,438	Sign in by email OTP; `auth.md`.
fund	2,294	Add funds through Coinbase Onramp; `fund.md`.
query-onchain-data	2,051	Query onchain data on Base through the CDP SQL API; `query-onchain.md`.
query-blockchain-data	1	No separate reference document in the tree; the nearest match is `query-onchain.md`, and the single install points to a duplicate listing rather than a second skill.

All 11 indexed listings appear in the table above.

The drift is the story here. One skill directory sits behind 11 listing names, and the README's capability table names nine reference files: auth.md, balance.md, send-usdc.md, trade.md, fund.md, x402-search.md, x402-pay.md, x402-monetize.md and query-onchain.md. Ten of the eleven listings map onto those files plus the router itself. The eleventh, query-blockchain-data, at one install, does not.

Why This Matters for Hermes Agents¶

Payment is the step where an autonomous agent either finishes the job or stops and asks a human, and this pack reduces that step to a CLI call the agent can plan around. Reading a balance, sending a stablecoin and settling a paid API call all become commands with a printed result, which is the shape a tool loop needs.

The layout is worth copying even if you never touch a chain. One SKILL.md with a routing table and a references/ folder per topic keeps the always-loaded context small and pushes the detail into files the agent opens only when the task calls for it. That is the difference between a skill that fits in a prompt and a skill that drowns it, and the README's contributing notes make the pattern explicit: new capability means a new reference file plus a new routing row, and a new shell tool means a new entry in allowed-tools.

The x402 half is the part with no obvious off-the-shelf substitute. An agent that can search a bazaar of paid endpoints, pay per call in USDC without a subscription, and stand up its own metered endpoint closes a loop that usually needs a human with a card. For a builder, that is the difference between an agent that consumes APIs and one that can sell access to its own.

Chain work also punishes optimism. A send needs a network, an amount and a recipient confirmed before anything is signed, and a swap needs a quote before it needs a signature. The balance and send references are what make a spend preflight possible, and the pinned CLI version means the flags an agent was tested against do not move underneath it.

Usage¶

You say	What happens
"Sign in to my wallet with me@example.com"	authenticate-wallet reads `auth.md` and starts the email OTP flow for the `awal` session.
"What is my USDC balance on Base and Polygon?"	The balance reference is read and per-chain balances are printed for the networks named.
"Send 10 USDC to barmstrong.eth"	send-usdc resolves the recipient, confirms chain and amount, then sends USDC.
"Swap 0.05 ETH for USDC on Base"	trade reads `trade.md`, obtains a quote and executes the swap on Base.
"Find a paid API for this and pay for one call"	search-for-service queries the x402 bazaar, then pay-for-service settles the call in USDC.
"Put my endpoint behind x402 at 0.01 USDC per call"	monetize-service builds and deploys the paid server from `x402-monetize.md`.
"How active was this contract on Base last week?"	query-onchain-data sends SQL to the CDP API through `query-onchain.md`.

Verification¶

The raw path below is built from the layout in this pack and returned 200 on branch main when this guide was written on 2026-11-03:

curl -sL -o /dev/null -w "%{http_code}\n" \
  "https://raw.githubusercontent.com/coinbase/agentic-wallet-skills/main/skills/agentic-wallet/SKILL.md"

After a clone, confirm that the directory count matches the pack and that the reference files the routing table names are present:

git clone --depth 1 https://github.com/coinbase/agentic-wallet-skills.git /tmp/cbx-skills
find /tmp/cbx-skills/skills -name SKILL.md | wc -l
ls /tmp/cbx-skills/skills/agentic-wallet/references
grep -o 'awal@[0-9][0-9.]*' -r /tmp/cbx-skills/skills | sort -u

The count should be 1, because the repository ships one skill directory. The listing should include auth.md, balance.md, send-usdc.md, trade.md, fund.md, x402-search.md, x402-pay.md, x402-monetize.md and query-onchain.md. The last command prints the pinned CLI versions, which should be a single version across every file if the bump script was run last.

Security¶

skills.sh verdicts for sampled skills (verified November 3, 2026) - re-check the security pages on skills.sh before production use:

Skill	Gen Agent Trust Hub	Socket	Snyk
trade	Pass	Warn	Warn
send-usdc	Pass	Warn	Warn
search-for-service	Pass	Pass	Warn
fund	Pass	Pass	Warn
query-blockchain-data	Fail	Pass	?

Only one of the five sampled listings passes all three auditors, and it is not a clean sweep: search-for-service and fund each carry a Snyk warning, trade and send-usdc carry both a Socket and a Snyk warning, and query-blockchain-data is the only listing in the repository with a Gen Agent Trust Hub failure. Its Snyk verdict is unresolved in this snapshot rather than passing.

The audit surface is a router plus nine Markdown references and one bump script, which is small enough to read in an afternoon. That is the mitigating fact and also the reason the verdicts deserve attention: the risk in this pack is not a sprawling dependency tree, it is that the instructions themselves move funds. Every reference that can sign a transaction should be read before the skill is given an agent with spending authority, and allowed-tools in the frontmatter is the first thing to review, because it decides which shell commands the skill is permitted to run.

Two domain risks do not show up in any scanner. Sign-in is an email OTP into a wallet session, so the mailbox is a key: whoever controls the address can approve the session. And a swap or a send is irreversible once broadcast, which means a misparsed amount is a real loss rather than a failed test.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations¶

Snapshot data, verified November 3, 2026: 27,257 combined installs across 11 indexed listings; 127 GitHub stars; license MIT; last pushed 2026-08-04. Counts drift over time.
One skill directory behind 11 indexed listing names. query-blockchain-data has a single install, no distinct reference document in the tree and the only Gen Agent Trust Hub failure in the sample, so treat it as a duplicate listing until the repository says otherwise.
The nine reference files carry the detail and the routing table decides which one is read. A vague request can land on the wrong reference, so name the chain and the amount when you ask.
The CLI version is pinned per code block rather than in a lockfile or manifest, which means an out-of-date pin is only visible by reading the files or running the bump script with Node v22 or newer.
Wallet access is an email OTP session rather than a scoped API key. There is no read-only mode documented in this snapshot, so the same sign-in that reads balances can send funds.
The repository does not document a Hermes install path, and its own examples assume the skills CLI. The profile copy above is the closest equivalent, not a supported route.

Skills Catalog | Skills Marketplace

Pro Tips¶

Read SKILL.md before you install anything else. The routing table is the whole product, and it tells you in one screen which capability costs a reference read and which tool each one is allowed to call.
Ask for a balance before a send or a swap. The balance reference exists, and a preflight read is cheaper than a failed transaction and a support thread.
Name the chain explicitly. USDC lives on Base, Polygon and Solana in this pack, and an unqualified send can resolve to a different network than the one you meant.
Pin the CLI and re-run the bump script deliberately. The version is written into every code block, so a partial bump leaves the agent mixing flag sets from two releases.
Keep query-blockchain-data out of your install list until the failure verdict is explained. query-onchain-data covers the documented onchain query path on its own.

← Skills Catalog | Skills Marketplace → Powered by CorpusIQ