# agentgrasp [Health: Active]

**Category:** 💻 Developer Tools  
**Repository:** https://github.com/damejeras/agentgrasp  
**GitHub Stars:** 0  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/agentgrasp

## Description
Yes/no questions about local files, and relevant-file search, with TypeSafe Jev

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "agentgrasp": {
    "command": "npx",
    "args": ["-y","agentgrasp"],
    "env": {
      "TYPESAFE_API_KEY": ""
    }
  }
}
```

**Requires environment variables:** `TYPESAFE_API_KEY` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What agentgrasp MCP server does

The agentgrasp MCP server changes how Claude Code handles Bash output. Each Bash command runs inside a grouped shell operation whose output is written to a state file. Claude Code receives a short summary and the path to that file rather than the complete output in its conversation context. The agent can then read the file directly when it needs the details.

The server also adds two file-analysis tools. `ask` answers yes/no questions about files within the configured MCP roots. `search` finds files and line ranges that relate to a question, while withholding the matching source text from its result. Both tools use Jev, TypeSafe's yes/no model.

## How it works

The Claude Code plugin invokes the `agentgrasp` executable from `PATH`. The executable intercepts the handling of Bash output, stores that output locally, and exposes the MCP functionality used for file questions and search. The project directory is an MCP root by default, and additional directories can be added.

When `ask` or `search` runs, file content is sent to TypeSafe for processing. Each call records the tokens it used beneath the agentgrasp state directory, in separate `asks` or `searches` locations. Search results identify relevant file locations and line ranges rather than returning source text.

## Setup and configuration

Install the release binary with the project's installer, or install the same binary through npm:

- `npm install -g agentgrasp`
- `npx agentgrasp mcp`

For Claude Code, add the `agentgrasp` plugin through its marketplace and install it. The environment that starts Claude Code must contain `TYPESAFE_API_KEY`; without this key, `ask` and `search` report `provider_unavailable`.

If Claude Code's sandbox is enabled, allow writes to `~/.local/state/agentgrasp`. When `XDG_STATE_HOME` is set, the corresponding directory is `$XDG_STATE_HOME/agentgrasp` instead. The state directory is not cleaned up automatically.

## Tools and capabilities

- `ask`: evaluates a yes/no question against local files.
- `search`: locates files and relevant line ranges for a question without returning source text.
- Bash output handling: saves long command output to files and gives Claude Code a summary plus the file path.

## Limitations and notes

The agentgrasp MCP server can read any file inside its MCP roots, including files that Claude Code `Read` deny rules block. Allowing `ask` or `search` therefore grants those tools access to the files in scope. Their requests send file content to TypeSafe, which lists Jev at $42 per billion input tokens.

Claude Code asks for approval for every Bash call containing agentgrasp's grouping wrapper in the default and `acceptEdits` permission modes, even when allow rules would otherwise permit the command. `bypassPermissions` runs those calls without a prompt, while deny rules continue to block commands. A runaway command can fill the disk until Claude Code stops it, and saved state files remain until removed manually.

_Full upstream README: https://allmcps.com/mcp/agentgrasp/readme_