[Feature]: Include Custom Tools in COMPOSIO_SEARCH

[jheising]

Area of Improvement

TypeScript / Node.js SDK

Feature Name

Custom Tools in COMPOSIO_SEARCH

Describe the Feature

Unless my understanding is incorrect, there isn't much gained by passing custom (memory-based) tools through the Composio client, any more than merging the tools together yourself right at the end, like:

tools: [...myTools, ...await session.tools()]

It'd be really handy if passing tools via the Composio client enabled those tools to work within the context of the COMPOSIO_SEARCH tool so that we could save on passing our tool context for every LLM turn.

Use Case / Motivation

Allows users to save LLM context for their custom toolsets by leveraging the power of the COMPOSIO_SEARCH tool.

Proposed Solution or API Design (Optional)

No response

Who does this impact?

• Developers using the SDK
• AI Agent frameworks (LangChain, OpenAI, etc.)
• Dashboard users
• End users of connected apps

Related Issues / References

No response

Would you like to contribute to this feature?

• Yes, I can help implement it
• Maybe, I can help test or review
• No, just sharing the idea

Additional Context

No response

[TensorVoyager]

"Hey @jheising 👋 I saw your Feature Request in Composio about COMPOSIO_SEARCH and saving LLM Context, and your idea is fantastic! Currently, the Agent needs to carry the full tool description with every poll, which consumes too much token.

As a developer also deeply optimizing the Agent toolchain, I'd like to share an experience from our data injection layer: besides dynamically loading tools at the registration layer, controlling the payload size returned by the Custom Tool is equally crucial for saving context.

Previously, when we had the Agent perform web searches or crawls, Raw HTML would often consume tens of thousands of tokens instantly. Later, our team specifically built the ThorData API as a Custom Tool for the Agent. It takes over the underlying anti-crawling and proxying, only converting the target webpage (or SERP results) into minimally structured JSON and sending it to the large model.

Combined with your envisioned on-demand loading logic, plus the underlying pure JSON data cleaning, the Agent's context consumption can be minimized. If you are building Custom Tools that require high-frequency network connectivity, you are very welcome to use our free API quota to run tests and see if it can help you." Think of how many tokens LLM saved! Looking forward to Composio supporting your PR soon! 🚀

[prasadf18]

hi @jheising please assign this to me