← Back

Technical documentation

SoloClaw is an autonomous agent that runs its own token on Solana. It observes the market, reasons about allocation, and executes actions on-chain — without human intervention.

Architecture overview

The agent operates in a continuous loop: observe → reason → act. Each cycle runs on a fixed interval and decides whether to claim fees, split capital, buy back tokens, burn supply, or add liquidity.

  1. Observe — Connects to Solana via RPC and reads on-chain state: creator vault balance, bonding curve status, pool depth, and market conditions.
  2. Reason — A reasoning layer evaluates whether to act. It can integrate with language models (e.g. Claude, Llama) to produce allocation strategies based on volume, fees, and migration state.
  3. Act — Builds and signs transactions using the Pump.fun and PumpSwap SDKs, then broadcasts them to the network.

SDK stack

The agent relies on official Solana and Pump.fun tooling. No custom RPC logic — everything goes through standard programs and SDKs.

  • @solana/web3.js — Connection to Solana RPC, transaction construction, signing.
  • @solana/spl-token — Token-2022 program for burn instructions, ATAs, transfers.
  • @pump-fun/pump-sdk — Creator fee collection, bonding curve state, buy instructions before migration.
  • @pump-fun/pump-swap-sdk — AMM swaps and LP deposits after the token has migrated off the bonding curve.

Execution flow

Each cycle follows this sequence. If any step fails or conditions aren't met, the agent skips and waits for the next cycle.

  1. Check vault balancegetCreatorVaultBalanceBothPrograms reads pending creator fees. Below threshold → skip.
  2. Split allocation — 80% to creator wallet, 20% to treasury. Uses SPL transfer instructions.
  3. Determine migration statusgetMinimumDistributableFee tells us if the token is on the bonding curve or migrated to Raydium.
  4. Buyback (bonding) — If on curve: PUMP_SDK.buyInstructions with Token-2022, then burn purchased tokens.
  5. Buyback + LP (migrated) — If on AMM: random split between buyback via buyQuoteInput and LP via depositInstructions. Burn bought tokens.

Token program

SoloClaw uses Token-2022 (SPL Token 2022 program) for all token operations: associated token accounts, buy instructions, and burns.

State persistence

Cycle results — claimed amounts, burns, LP additions — are written to a database. The website reads this state to display the “thought” and feed. The agent does not store secrets or keys in the database; only aggregated stats and the latest thought string.

Deployment modes

The agent can run in two modes:

  • Serverless — Triggered by site visits or a cron job. Runs in a serverless function, signs with a configured keypair.
  • Local + tunnel — Runs on your machine, keys never leave. Exposed via a tunnel (e.g. ngrok) so the website can trigger cycles remotely.

Links