Codebase Agent MCP

What FI-Mihej/codebase-agent-mcp MCP server does

FI-Mihej/codebase-agent-mcp MCP server separates repository exploration from the primary coding agent. A dedicated OpenAI-compatible model analyzes source files, dependencies, documentation, and other large information sets, then sends back findings relevant to the active task. This can reduce the amount of repository content that the primary model must repeatedly read and retain in context.

The project is aimed at the 🤖 Coding Agents category, but its stated use extends beyond software repositories. The author describes using it for documents, scientific papers, and large collections of poorly searchable web content. It can therefore fit long-horizon research or analysis tasks when the connected model has a smaller context window or lower operating cost.

How it works

Connect a local or low-cost cloud model with an OpenAI-compatible API to the server. The server can use local inference systems such as LM Studio, llama.cpp, vLLM, and SGLang, or compatible hosted services such as OpenRouter. Then connect the server to an MCP-compatible harness, including Claude Code or Codex. The README recommends beginning a development prompt with Use skill libraries-analysis-skill. to reduce costs.

The subharness manages conversation-history compaction, configurable recursive subagent depth, task-completion control, and an anti-stuck mechanism for repeated generation. It can also run asynchronous background jobs and persist their results in SQLite. External stdio MCP tools are supported, allowing the analysis process to use connected MCP capabilities.

Setup and configuration

The material confirms two transport modes: STDIO and HTTP. It does not provide installation commands, a complete configuration example, or a documented list of environment variables in the supplied excerpt. Model access must use an OpenAI-compatible interface, and filesystem access is sandboxed to configured library roots.

Optional Qdrant integration is available through built-in qdrant_* plugins. The current use described for this integration is caching retrieval results; the project plans to expand it toward semantic retrieval of code entities such as files, classes, and functions.

Tools and capabilities

• Analyze codebases and dependencies without sending every file repeatedly through the primary agent.
• Delegate documentation and broader information-retrieval work to a separate model.
• Use local models or compatible hosted APIs.
• Connect external stdio MCP tools to the subharness.
• Run recursive subagents with configurable depth.
• Persist asynchronous job results in SQLite.
• Use optional Qdrant-backed retrieval caching.

Limitations and notes

FI-Mihej/codebase-agent-mcp MCP server depends on an OpenAI-compatible chat-completions interface rather than a framework such as LangChain. The excerpt does not document a fixed model list, installation procedure, environment-variable configuration, or named MCP tool schema. Qdrant semantic retrieval of code entities is described as planned rather than an established capability.

The benchmark reports lower token use and estimated costs for one Cengal analysis scenario, with VectorDB caching disabled for the test. Those measurements are examples from the project and should not be treated as a guarantee for every repository, model, or workload.