A non-custodial trading daemon steered over MCP. An AI agent host sets standing price rules; the daemon watches prices and acts on them by itself. The agent has no tool that quotes, signs or broadcasts a swap, so a rule is the only thing it can commit. See docs/SKILL.md for the agent contract.
Execution runs on Uniswap through its Trading API, and on Cetus for Sui CLMM pools when sui.enabled is set. One loop drives both: a venue builds, the vault signs, the chain client broadcasts. See ARCHITECTURE.md for architecture details.
tempo-agentic-daemon is the binary that trades. Broadcasting is off unless MAINNET_SWAP=1 is set, so run quotes, builds and signs without sending anything. On Sui such a run also asks the node to simulate the transaction it built, so a run that spends nothing still says whether the swap would have worked; the answer lands in the order's failure reason.
Signing keys live in the files keys.evm and keys.sui point at, one raw key per file, owner-readable only. bootstrap creates whatever is missing; keystore import asks for an existing key without echoing it.
tempo-agentic-daemon bootstrap # create accounts; initialize the state DB if absent
tempo-agentic-daemon keystore import --chain evm
tempo-agentic-daemon health # also prints the addresses it will sign with
tempo-agentic-daemon strategy add --id eth-usdc --chain base \
--base-token WETH --quote-token USDC
tempo-agentic-daemon strategy list
tempo-agentic-daemon level add --id buy-3000 --strategy-id eth-usdc \
--side buy --trigger-price-usd 3000 --amount 2 --slippage-bps 50
tempo-agentic-daemon level list
tempo-agentic-daemon level rm --id buy-3000
tempo-agentic-daemon run
tempo-agentic-daemon dashboard # run in another terminal after `run`
tempo-agentic-daemon resolve-quarantine --order-id o-1A strategy owns its market. Every level refers to one with --strategy-id: buy spends the strategy's quote token for its base token (quote -> base), while sell spends base for quote (base -> quote). --amount is expressed in whole units of the token being spent.
bootstrap and run initialize the current schema only when the state database does not exist. There is no schema migration: if an older development database is rejected, remove the exact path named in the error manually and run bootstrap or run again. The daemon never deletes it for you.
strategy add, level add, and level rm write straight into SQLite when nothing holds it. When run holds the database lock they call the daemon's own MCP tools instead, so a trading daemon never has to stop to be reconfigured. The daemon validates whatever it is asked to store, so the checks are the same either way.
openclaw prints the mcp.servers.nofomo entry for the running daemon, and --write merges it into ~/.openclaw/openclaw.json without touching anything else in that file:
tempo-agentic-daemon openclaw
tempo-agentic-daemon openclaw --writeThe entry carries a toolFilter that exposes only authoring: set_strategy and set_level, plus strategies, levels and status so an agent can see what already exists and whether a rule it stores will trade for real. delete_level and orders stay out. Widen the list in openclaw.rs if an agent should do more than write rules.
The daemon publishes a fresh port and token every start, so run this again after restarting it. The written file holds a bearer token and is kept owner-readable.
dashboard requires a running daemon. It reconnects through the daemon manifest after connection or authentication failures and renders feed and level states as text (for example live, stale, armed, or cooldown). Press q or Ctrl-C to detach; trading continues in the daemon.
A price belongs to the asset, not to the chain it sits on — arbitrage keeps them level. EVM tokens are quoted by their own address, while each entry in sui.coins names a price_ref: the same asset on a chain the feed does index. That is how a testnet coin no feed has ever heard of still gets a price. A coin used only as a strategy's quote leg can omit it, but then it cannot be the base token a strategy is watched on.
Before an order is created, the venue's quote is checked against the price that fired the level and refused if it sits further away than max_quote_deviation_bps (5% by default). This catches what slippage cannot: minimum_amount_out is derived from the venue's own quote, so a bad route or a manipulated pool produces a self-consistent quote the daemon would otherwise sign. The check values the counter leg in dollars, so it needs usd_peg: true on the quote token — mark your stablecoins. An unpegged pair is logged as unchecked rather than silently trusted.
Strategies on Sui name coins by the symbols in sui.coins and need --venue cetus:
tempo-agentic-daemon strategy add --id hbtc-sui --venue cetus --chain sui \
--base-token hBTC --quote-token SUI
tempo-agentic-daemon level add --id sell-100k --strategy-id hbtc-sui \
--side sell --trigger-price-usd 100000 --amount 0.001 --slippage-bps 100Cetus needs a pool with real liquidity behind the pair. create-pool opens one and seeds it, signing with the same vault key the daemon trades from:
create-pool --config infra/local/config.json \
--coin-a hBTC --coin-b SUI \
--amount-a 100000 --amount-b 1000000000 --price 100000 --dry-run--dry-run resolves the coins, reads their decimals on chain and prices the transaction without sending it. The registry defaults to the one in crates/cetus/src/constants.rs, which is the one this venue looks pools up in — Cetus runs more than one deployment on testnet, and a pool opened against another registry is one the daemon will never find. --pools overrides it only if Cetus redeploys.
cargo fmt --all -- --check
cargo clippy --all-targets --all-features -- -D warnings
cargo test --workspace --all-targets --all-features
make buildapps/ contains binaries. crates/ contains library components. docs/ contains MCP integration guides. infra/ contains local development infrastructure.