Clawmes

Clawmes turns any Hermes Agent (Nous Research) into a full on-chain operator you

drive in plain English. Tell it what you want — "watch the six wallets I follow

on Base, and if any buys over $50k of a token with $100k+ liquidity, snipe 0.15

ETH with MEV protection, set a +100% take-profit and a -30% stop" — and it

parses that into a concrete, running rule. Natural language is the interface;

there are no forms, scripts, or dashboards to wire up.

It's built for conditional and multi-step operations, not one-shot trades:

threshold-gated rebalances ("rebalance to 50/30/20, but only if a leg drifts

more than 5%"), two-sided price ladders ("if ETH breaks $4,000 sell 25% of my

stETH; under $3,000 buy it back"), filtered copy-trading, and full bridge → swap

→ LP flows — each compiled into triggers, brackets, and schedules that run on

their own and report back.

Safety is cryptographically enforced, not just promised. Rather than handing

over keys, you grant the agent scoped, revocable on-chain permissions through

MetaMask Smart Accounts — ERC-7715 permission grants and ERC-7710 delegation

with caveats on ERC-7702-upgraded accounts — so it can only ever do what you

authorized: spend caps, time limits, allow-listed actions, enforced at the

wallet. On top of that, fund-moving actions still require explicit confirmation

with full details shown, the agent verifies the live wallet/chain read-only

before acting and never fabricates results, swaps route through private

orderflow, launches/snipes run honeypot + tax pre-checks, and approvals can be

audited and revoked.

It's open at every layer. Wallets: MetaMask (WalletConnect v2 + MetaMask Smart

Accounts), plus Bankr, Coinbase, Rabby, or a local key across Ethereum, Base,

Arbitrum, Optimism and Polygon. Inference: pluggable OpenAI-compatible providers

for the agent's own research and summarization — Venice (privacy-first) or

OpenGateway, selectable per deployment. Agent-to-agent: discover and hire other

agents over the A2A protocol (JSON-RPC 2.0 AgentCards) and the Clawnx matching

network, delegating sub-tasks and running agent-to-agent commerce. Machine

payments: the agent pays x402-gated APIs and inference per request from its own

wallet in USDC on Base — no API keys, no checkout, fully autonomous.

Underneath: 53 tools — 0x swaps, LiFi bridging, Aave lending, Lido / Rocket Pool

staking, Uniswap V3 liquidity, Reservoir NFTs, Clawnch + Bankr launches,

prediction markets, Snapshot/Tally governance, Gnosis Safe, Permit2, and an

automation suite (DCA, copy-trading, bracket limit orders, sniper, alerts).

One-line install (pip install clawmes), rich cards in Hermes Desktop, and

production-grade engineering — ~4,600 tests at 100% line coverage. It plugs into

the $CLAWNCH ecosystem and the Clawnch launchpad on Base.

Over the hackathon (May 18 → June 9) clawmes went from v0.15 to v0.18 — a dozen+

merged PRs across the plugin, the Clawnch launchpad, and the clawncher CLI/SDK,

holding 100% line coverage the entire way. Highlights:

• Natural-language automation (v0.15): shipped the trader suite — MEV-protected

swaps, bracket limit orders, wallet scan, airdrop eligibility/claim — and the

conditional-rule engine behind snipe / rebalance / price-ladder / filtered

copy-trade ("if a wallet I follow buys >$50k… snipe with a +100/-30 bracket").

• Hermes Desktop UI (v0.16.x): built the card system — portfolio, research,

launch-receipt, and WalletConnect-QR cards — with explorer/Dexscreener link

enrichment across every transaction tool and click-to-open previews. Traced

and fixed the desktop preview-envelope mechanism against the live gateway.

• Agent-callable surface (v0.17.0): added clawmes_info so the agent can pull

wallet/balance/portfolio/research/trending/leaderboard data itself (53rd tool).

• Wallets & on-chain guardrails: hardened WalletConnect (env-only project ID),

and wired MetaMask Smart Accounts via the sa-bridge — the ERC-7702/7710/7715

stack for scoped, revocable agent permissions (spend caps, time limits,

allow-listed actions enforced at the wallet, instead of handing over keys).

• Inference providers (v0.18.0–0.18.1, + 0.18.2 in progress): added Venice

(privacy-first, OpenAI-compatible) alongside OpenGateway behind a unified

provider router, live-tested against the real Venice API, and handled the

x402 (HTTP 402) machine-payment path — the agent can pay per-request APIs and

inference in USDC on Base, no keys or checkout.

• Launchpad (Clawnch): shipped mandatory 1M $CLAWNCH burn enforcement across all

launch paths (custodial /api/deploy, non-custodial /api/prepare, and the

4claw/Moltbook cron scanners), documented the launch requirements, and trimmed

infra cost.

• clawncher CLI/SDK: built and published to npm the client-side burn flow

(clawncher deploy --burn, SDK burnClawnch) so any agent can satisfy the launch

burn and deploy a verified token in one command.

Engineering throughout: conventional commits, CI-gated 100% coverage (~4,600

tests), and automated PyPI/npm publishing on tag.

Bootstrapped