> For the complete documentation index, see [llms.txt](https://docs.thetanuts.finance/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.thetanuts.finance/sdk/ai-agents/agents-overview.md).

# Overview — Pick Your Route

Thetanuts ships two agent-facing packages on top of the SDK. Both are thin layers over the same [`@thetanuts-finance/thetanuts-client`](https://www.npmjs.com/package/@thetanuts-finance/thetanuts-client) encode helpers — the difference is **who signs the transaction and where the safety boundary lives**.

## The two packages

|                 | [`@thetanuts-finance/mcp`](https://www.npmjs.com/package/@thetanuts-finance/mcp)                                    | [`@thetanuts-finance/agentkit`](https://www.npmjs.com/package/@thetanuts-finance/agentkit) |
| --------------- | ------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------ |
| What it is      | MCP server: \~100 tools an LLM client calls over the Model Context Protocol                                         | Coinbase AgentKit `ActionProvider` library you embed in your own agent code                |
| Who signs       | **Never signs.** Pairs with a signer — Base MCP (you approve each tx in Base Account), Safe, or a CDP policy wallet | **The agent's own wallet** (CDP, viem, Privy server wallets), unattended                   |
| Runs in         | Claude Desktop, Claude Code, Cursor, ChatGPT, Codex — any MCP client                                                | Your backend agent process (LangChain, Vercel AI SDK), or as its own MCP server            |
| Safety boundary | Outside the LLM: wallet approval UI or signer policy                                                                | In code: fail-closed `SafetyPolicy` (notional caps, collateral allowlist, host hook)       |
| Use when        | Human-in-the-loop chat trading; maximum client reach                                                                | Headless trading bots, MM bots, custodied agent vaults                                     |

This split is deliberate: the MCP server can guarantee it **cannot move funds** (it holds no keys and builds calldata only), while autonomous signing stays an explicit, separately-installed opt-in. The two can run side by side.

## The three routes

| You want                                                  | You run                                                                                                                                     | Guide                                                |
| --------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------- |
| **Trade from chat, approving every transaction yourself** | `@thetanuts-finance/mcp` + [Base MCP](https://docs.base.org/ai-agents/quickstart) — calldata is prepared, you click approve in Base Account | [Trade from Chat](/sdk/ai-agents/base-mcp-plugin.md) |
| **Trade from chat, but the agent signs by itself**        | `@thetanuts-finance/agentkit` run as an MCP server (Coinbase's MCP adapter + a CDP wallet under `SafetyPolicy` caps)                        | [AgentKit](/sdk/ai-agents/agentkit.md)               |
| **A fully headless bot — no chat client at all**          | `@thetanuts-finance/agentkit` embedded in your own code (LangChain, Vercel AI SDK)                                                          | [AgentKit](/sdk/ai-agents/agentkit.md)               |

If you're unsure, take the first route — a transaction can never leave your wallet without your click.

## Reads only?

If you just want an LLM that can *read* the protocol (markets, positions, IV, Greeks) and help you write SDK code, the MCP server alone is enough — no wallet, no signer. See [MCP Server](/sdk/ai-agents/mcp-server.md), or skip servers entirely with the copy-paste [LLM Context](/sdk/resources/llm-context.md) prompt.

## Deeper material

* [MCP Server guide](/sdk/ai-agents/mcp-server.md) — every tool, environment variables, the full comparison
* [Trade from Chat](/sdk/ai-agents/base-mcp-plugin.md) — the Base MCP plugin setup
* [AgentKit guide](/sdk/ai-agents/agentkit.md) — autonomous agents, both modes
* AgentKit repo: [SETUP.md](https://github.com/Thetanuts-Finance/thetanuts-agentkit/blob/main/SETUP.md) (end-to-end setup: CDP wallet, client configs for Claude Desktop/Code, Cursor, Codex) and [SKILL.md](https://github.com/Thetanuts-Finance/thetanuts-agentkit/blob/main/SKILL.md) (a walkthrough skill any skill-aware agent can load)


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://docs.thetanuts.finance/sdk/ai-agents/agents-overview.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.