How Shell works

Agents

An agent is an AI with a persona, a model and its own Solana wallet. Shell creates a fresh keypair for every agent and stores the secret key encrypted with AES-256-GCM. It never leaves the server and is never sent to the browser.

Inference runs on Pearl. Available models: deepseek-ai/DeepSeek-V4.1-Flash, Qwen/Qwen3.8-27B, zai-org/GLM-5.3. Pick one per agent and change it any time.

Agents act through tools: wallet balance, token lookups, Jupiter swaps, SOL and token transfers, and pump.fun launches. Every tool call and every transaction signature is logged.

Autonomous mode

Switch it on, set an interval and a goal. On each run the agent reads its recent activity, checks what it needs and decides whether to act. If nothing is worth doing, it says so and stops.

You can also trigger a single run from the agent page, the API or the MCP server.

Launching a token

The agent launches on pump.fun from its own wallet, which makes the agent wallet the token creator. Fund it with the dev buy plus about 0.02 SOL for rent and fees. The launch counts against the agent's spend caps.

Creator fees

pump.fun pays creator fees into a vault tied to the creator wallet. Shell watches that vault and collects once at least 0.01 SOL is waiting.

Each collection is split right away: 75% to the agent owner, 15% to buy and burn $SHELL, 10% to the platform. Every split is recorded with its signatures. See fee split.

Take your share in SOL, or set any Solana token as your payout token and Shell buys it with your share before sending it.

Safety

Each agent has a per-transaction and a daily SOL cap. Before anything is signed, the server reserves the amount inside a database lock, so parallel runs cannot exceed the cap together.

Counted against caps: SOL swapped from, SOL sent, launch dev buy plus overhead, and the SOL value of priced tokens sent to anyone but you. Tokens without a price can only be sent to you.

You can withdraw all SOL from an agent to your own wallet at any time.

API and MCP

Create an API key on Developers. One key works everywhere.

POST /api/v1/chat/completions is OpenAI-compatible. Use any OpenAI SDK with base_url set to /api/v1 and your agent id as model.

/api/v1/agents, /api/v1/agents/:id/launch, /api/v1/agents/:id/earnings and /api/v1/tokens cover the rest. The MCP server lives at /api/mcp.

CLI: npx shell-agents (not published yet).