This is a fork to implement a Tool Search Tool for OpenCode, mimicking Claude's Tool Search Tool . Such a tool allows for two main things:

• Keeping context free of tool descriptions and schemas. Especially helpful for local LLM users that seek to save VRAM by keeping context windows small
• Give agents up to 10000 tools to work with, with normal context usage. Potentially enables agents with unprecendented power and ability

For the tool search tool, we built a Catalog system, where each session can add CatalogEntrys to its own tool set. Each CatalogEntry corresponds to an available tool (built-in or MCP) that is either loaded by default, or has deferLoading set to true. Built-in and plugin tools are loaded via the native ToolRegistry.all() getter-function, and the MCP tools come from MCP.toolsMeta(). All the tools are deferred for loading by default, except for the tool search tools (set via the DEFAULT_CORE_TOOLS flag internally), but this can be configured in the your opencode installation's config directory (see "How To Run").

There are two versions of the tool search tool loaded by default: BM25 tool search and RegEx tool search, in the spirit of Claude's existing implementation.

The BM25 tool search tool is loaded as the tool_search tool and resides in (src/tool/tool-search.ts). Given a query, it uses BM25's ranking technique to search for tools that have a (case-insensitive) keyword matching in their search. For more information look up the Wikipedia page that details how the potential results of a query get scored/ranked.

We set the default parameters to k1 = 0.9 and b = 0.4. Both are relatively low values (from common middle-term values of k1 = 1.5 and b = 0.75), with the rationale that a low k1 value doesn't emphasize the frequency of the searched term in the query that much, encouraging "semantic density". The low b value is chosen such that longer tool descriptions aren't agressively normalized and hidden when searching.

These values were specifically chosen for Small Language Models (SLMs) that tend to send vaguer descriptions of the tools they want to search for. If the model you are working with is highly capable, and writes very specific queries for large sets of tools, we encourage you to increase the k1 value and play around with the b value such that the results fit your context efficiency needs.

As an alternative to BM25, we have the Regex tool search tool, that allows the agent to search for tools with a regular expression (case-insensitive). It returns all the matches to the names and descriptions of the tools. This is for queries where agents have a more specific idea of what they are searching for, and differs from BM25 in the sense that not term frequency but term composition is important.

Currently, two tools are load, allowing the agent to search for a tool and its description. For more information on how BM25 works check out the

1. Tools, that are always loaded into context

In ~/.config/opencode/opencode.json add the following lines:

    "toolSearch": {
      "alwaysLoad": ["read", "write", "edit"]
    },

This allows the agent to be aware about its ability to read, write and edit a file from the very beginning.

2. Start the development server to use this forked version of OpenCode

bun run dev

We have two simple CLI scripts (RegEx & BM25) for you to test them with your own inputs. Run the following commands:

cd packages/opencode
bun run src/search/test-bm25-cli.ts # To Test BM25 implementation
bun run src/search/test-bm25-cli.ts # To Test RegEx implementation