Agent Lsp

What blackwell-systems/agent-lsp MCP server does

The blackwell-systems/agent-lsp MCP server acts as an orchestration layer between AI agents and existing Language Server Protocol implementations. It is not itself a language server. Instead, it manages servers such as gopls, rust-analyzer, and jdtls, then exposes their code-intelligence operations through MCP.

The README describes a single Go binary with 65 tools, 31 CI-verified languages, and 24 agent workflows. Supported operations include finding symbols and references, inspecting callers and callees, retrieving diagnostics, requesting formatting edits, getting completions, and examining semantic tokens or inlay hints. Workspace folders can be added for multi-repository analysis, including references from a shared library into consumer projects.

How it works

A persistent runtime keeps language-server indexes available across agent sessions. The first session may spend time indexing a workspace; later requests can reuse that state. start_lsp must be called before operations such as adding workspace folders or exporting the reference cache.

For safer modifications, an agent can create a simulation session, make in-memory changes, and compare diagnostics with the baseline. It can then commit the session with changes applied to disk, request a patch without applying it, or discard the session. The documented workflow places preview and impact analysis before applying edits. Tools such as blast_radius, find_callers, find_references, and detect_changes help identify affected code and distinguish test callers from non-test callers.

The server routes files to language-specific servers based on the project and configured environment. detect_lsp_servers scans a workspace for source languages and checks whether corresponding server binaries are available on PATH.

Setup and configuration

The README provides this installation command:

curl -fsSL https://raw.githubusercontent.com/blackwell-systems/agent-lsp/main/install.sh | sh && agent-lsp init

After installation, agent-lsp init is the documented initialization step. The repository describes the executable as a single Go binary. Because the project orchestrates external language servers, the available language-server binaries and their locations matter; detect_lsp_servers can report detected languages, installed servers, and a suggested configuration.

Responses use GCF, or Graph Compact Format, by default. Set AGENT_LSP_OUTPUT_FORMAT=json to switch the output format back to JSON.

Tools and capabilities

The blackwell-systems/agent-lsp MCP server includes tools for:

• Searching symbols, references, callers, completions, highlights, inlay hints, and semantic tokens.
• Reading diagnostics and requesting document or range formatting edits.
• Previewing, applying, simulating, evaluating, committing, or discarding edits.
• Measuring change impact through blast-radius and git-change analysis.
• Managing workspace folders, open documents, LSP notifications, and cached references.
• Enforcing ordered phases for skill workflows with activation and deactivation tools.

Limitations and notes

Results depend on the language servers available to the runtime and the capabilities those servers implement. For example, inlay hints and semantic tokens return useful results only when the selected language server supports the relevant LSP method. The supplied material does not specify configuration examples for every language, MCP client compatibility details, or a hosted endpoint.

The blackwell-systems/agent-lsp MCP server is licensed under MIT according to the repository badges. It is intended for local or CI-oriented code intelligence workflows rather than database access, despite the current directory category shown in the listing.