Jev agent tools

What Jev agent tools does

The Jev agent tools MCP server connects an MCP client to a Jev API-format endpoint and a local repository. It provides six tools for investigating unfamiliar code and evaluating changes:

• jev_ask combines a note, selected files, earlier versions, or command output into one evidence-based judgment.
• jev_ask_files applies the same question independently to multiple candidate files.
• jev_find_files identifies likely entry points for a behavioral goal when the filename is unknown.
• jev_locate_in_file finds a relevant range inside a single file of at least 19 KB.
• jev_check_diff reviews completed changes for risk, existing documentation, or specification concerns.
• jev_select_tests suggests commands for affected existing tests without running or collecting them.

These tools complement native file reading, search, and execution. They are not a replacement for exact source inspection, known-symbol lookup, line counting, arithmetic, or running commands yourself.

How it works

The MCP package starts as a stdio process and exposes the same six tools available through the pi and omp integrations. Each process represents one session. The server uses a configured Jev endpoint, bearer credential, requested model, and repository root. JEV_TOOLS_ROOT supplies the root when the MCP launch configuration does not pass --root.

Requests can be limited with session-wide call and cost ceilings. The jev_ask schema can also include command execution unless JEV_TOOLS_ALLOW_COMMAND=0 is set. Enabled commands use ordinary shell permissions and do not receive the API key. On Windows, command evidence uses Git for Windows Bash unless another executable is supplied through JEV_TOOLS_BASH.

Setup and configuration

Install the npm package and launch its MCP entry point with Node.js 24 or later. A typical client configuration runs npx -y -p jev-agent-tools jev-agent-tools-mcp --root /path/to/repository. Windows clients generally need cmd /c npx at the beginning of the argument list.

Set JEV_TOOLS_URL to a complete HTTPS-compatible Jev endpoint and JEV_TOOLS_API_KEY to the required Bearer credential. Plain HTTP is accepted only for localhost addresses. JEV_TOOLS_MODEL defaults to openjev; JEV_TOOLS_MAX_CALLS and JEV_TOOLS_MAX_USD add optional session limits. The server can also read saved configuration from the user configuration directory, after environment variables. Stored keys are plaintext, although the configuration file and directory receive restricted permissions.

Tools and capabilities

For developers evaluating fit, the Jev agent tools MCP server is most relevant when an agent must explain a repository-backed judgment rather than produce an unsupported answer. File discovery and in-file location help narrow evidence before asking a question. Diff review focuses on completed changes, while test selection identifies existing affected tests but does not execute them.

The package is documented for pi, omp, and MCP clients including Claude Code, Claude Desktop, Cursor, VS Code, and Codex CLI. It is also listed in the official MCP Registry under io.github.NomenAK/jev-agent-tools.

Limitations and notes

The endpoint and API key are required for useful tool requests. Without them, tools remain registered but report missing configuration, and there is no fallback to a chat model. Optional native parsers and accelerated file search may be unavailable on some platforms; affected tools report those limitations.

Git is required, and Bash is needed for optional command evidence. The automatic run-end documentation check can be disabled with JEV_TOOLS_AUTO_DOCS=0; MCP does not provide an automatic run-end documentation hook. The published setup also notes a pi packaging warning involving duplicate host-provided TypeBox modules, although the documented smoke test completed successfully.

The server requires Git and, for optional command evidence, Bash. The repository recommends using native reading, searching, and execution for exact facts, known locations, counts, arithmetic, and command results.