Skip to main content
Glama
Chromia LSP MCP

Chromia LSP MCP

Official

Rell LSP MCP Server

Rell language intelligence for AI coding agents, over the Model Context Protocol. It ships as a container image with the Rell language server inside, so there is nothing to install beyond a container runtime.

The canonical repository is on GitLab: gitlab.com/chromaway/core-tools/chromia-lsp-mcp. The image is published to that project's container registry. The old npm package @chromia/chromia-lsp-mcp is no longer maintained.

What it gives an agent

Agents like Claude Code and Cursor only see Rell as text. This server gives them what an IDE has:

  • Types, docs, and the meaning of the symbol at any position

  • Completions actually in scope there

  • The compiler's and linter's own errors and warnings

  • Navigation: definitions, references, outlines, project-wide symbol search

  • Edits written to disk: rename, quick fixes, formatting

Related MCP server: LSP MCP

Requirements

Only an OCI compatible container runtime.

Images are published for linux/amd64 and linux/arm64, so Apple Silicon runs natively.

Setup

An MCP entry is bound to one project: the project directory appears twice in the run command, once as the mount and once as the working directory. Configure this server per project, not once globally — a single global entry cannot follow you from one dapp to the next, and aiming one at a parent directory holding several projects makes the language server index all of them at once. Every client below is therefore configured inside the project it serves.

Mount the project root — the directory holding chromia.yml and src/ — because only what is mounted is visible to the language server.

Pick your runtime

On Linux, use docker with --user "$(id -u):$(id -g)" — without it, files the server edits come back owned by root. Podman needs no such flag.

On macOS, use Apple's container, started once per boot with container system start, or Docker Desktop or Colima with docker. No --user needed. Docker Desktop shares only /Users by default.

On Windows, work in WSL2 with Docker Desktop's WSL integration and follow the Linux line, keeping the project on the WSL filesystem rather than /mnt/c/.... Windows-native Docker cannot work: a Linux container has no C:\... path.

Examples below use docker; substitute container or podman.

Claude Code

Run this from the project root. The shell expands $(pwd) as the entry is written, so it is pinned to this project. On Linux:

claude mcp add chromia-lsp -- docker run --rm -i \
  --user "$(id -u):$(id -g)" \
  -v "$(pwd):$(pwd)" -w "$(pwd)" \
  registry.gitlab.com/chromaway/core-tools/chromia-lsp-mcp:latest

On macOS, with Apple's runtime started (container system start):

claude mcp add chromia-lsp -- container run --rm -i \
  -v "$(pwd):$(pwd)" -w "$(pwd)" \
  registry.gitlab.com/chromaway/core-tools/chromia-lsp-mcp:latest

claude mcp add defaults to local scope, which is what you want here: the entry is stored under this project's path and loads only in it. Avoid --scope user, which would offer one project's hardcoded path in every other project you open. claude mcp list shows whether it connects, and claude mcp remove chromia-lsp undoes it.

For a team, --scope project writes the entry to .mcp.json in the repository instead. That file travels to other machines, so the absolute path has to come from somewhere per-developer — Claude Code expands ${VAR} in arguments, so have everyone set one variable in their shell:

"args": [
  "run", "--rm", "-i",
  "-v", "${RELL_PROJECT_DIR}:${RELL_PROJECT_DIR}",
  "-w", "${RELL_PROJECT_DIR}",
  "registry.gitlab.com/chromaway/core-tools/chromia-lsp-mcp:latest"
]

Cursor

Cursor keeps project servers in .cursor/mcp.json at the project root. It substitutes ${workspaceFolder}, so the file resolves to the right path on every machine and can be committed:

{
  "mcpServers": {
    "chromia-lsp": {
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "-v", "${workspaceFolder}:${workspaceFolder}",
        "-w", "${workspaceFolder}",
        "registry.gitlab.com/chromaway/core-tools/chromia-lsp-mcp:latest"
      ]
    }
  }
}

~/.cursor/mcp.json is the global equivalent and the wrong home for this server: one entry cannot serve two projects.

Other MCP clients

Copilot in VS Code reads .vscode/mcp.json in the project, names the key servers instead of mcpServers, and substitutes ${workspaceFolder} the same way Cursor does. The entry is otherwise identical.

A client with no variable substitution needs the project's absolute path spelled out:

{
  "mcpServers": {
    "chromia-lsp": {
      "type": "stdio",
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "-v", "/home/you/my-dapp:/home/you/my-dapp",
        "-w", "/home/you/my-dapp",
        "registry.gitlab.com/chromaway/core-tools/chromia-lsp-mcp:latest"
      ]
    }
  }
}

Keep that file out of version control — the path in it is true only on your machine.

Pinning a version

:latest follows releases. Pin a specific one when you want the language server version to stay put; each image tag carries one Rell language server, recorded in the image label com.chromia.rell-lsp.version. Image tags match the repository's git tags.

Using it

Your MCP client starts the container; you never run it by hand. Ask the AI agent to work on Rell code and it drives the tools itself, starting with start_lsp. That one defaults to the mounted working directory, so in the setups above it needs no argument — if the agent ever picks the wrong place, name the right one:

Start the Rell LSP server with root directory /home/you/my-dapp

The container lives as long as the client session and holds the language server's index in memory, so the first query after startup is the slow one.

Tools

Tool

What it does

start_lsp

Start the language server on a project root. Call this first.

restart_lsp_server

Restart it, optionally on a different root

open_document

Open a file for analysis

save_document

Push the file's current content so diagnostics refresh

close_document

Close a file and release what it holds

get_diagnostics

Errors and warnings for one file or all open files

get_info_on_location

Hover: types, docs, and context at a position

get_completions

Completions valid at a position

get_code_actions

Quick fixes and refactorings for a range

apply_code_action

Apply one of them, writing the result to disk

get_definition

Where a symbol is declared

get_references

Every use of a symbol

get_document_symbols

Outline of one file

get_workspace_symbols

Find a symbol anywhere in the project

rename_symbol

Rename across every file that references it

format_document

Format a file or a range within it

set_log_level

Change logging verbosity at runtime

Line and column arguments are 1-based, the way an editor reports them.

Resources

  • lsp-diagnostics:// — diagnostics for every open file, and lsp-diagnostics:///path/to/file.rell for one of them. Subscribe to either and the server notifies you whenever the language server republishes.

  • lsp-hover:///path/to/file.rell?line=6&column=8

  • lsp-completions:///path/to/file.rell?line=25&column=10

Logging

The server reports what it is doing as MCP log notifications. Ask the assistant to set the log level to debug, or start it that way:

-e LOG_LEVEL=debug

added to the docker run arguments. In Claude Code, claude --mcp-debug additionally shows the raw traffic between client and server. Everything the server writes for humans goes to stderr; stdout carries the protocol and nothing else.

Environment variables

Variable

Default

Purpose

LOG_LEVEL

info

debug, info, notice, warning, error, critical, alert, emergency

RELL_LSP_JAR

/opt/rell-lsp/language-server.jar

Language server to run

RELL_LSP_JAVA_OPTS

—

JVM flags for the language server process, e.g. -Xmx2g

How the mount works

The server runs inside a container, but the agent talks about files by path — the paths it sees on your machine. So the project is mounted at the same path inside the container:

-v "/home/you/my-dapp:/home/you/my-dapp" -w "/home/you/my-dapp"

With that, /home/you/my-dapp/src/main.rell means the same file on both sides and no translation is needed anywhere. Every example in this README follows that pattern.

Troubleshooting

Cannot read /home/you/my-dapp/src/main.rell means the file is outside the mount. Check that the -v path is the project root and that -w matches it.

Empty diagnostics are usually a timing problem: the language server pushes them after indexing, and only for open files. Open the file first and retry; on a large project the first pass takes a few seconds.

Files coming back owned by root on Linux means the --user "$(id -u):$(id -g)" flag is missing from the run arguments.

If every tool returns an empty result, the language server probably failed to start. Set LOG_LEVEL=debug and read the client's MCP server log — the language server's own output is forwarded there.

Security

An MCP server runs on your machine with your files, so what it executes matters. The retired npm package downloaded the language server — and on most platforms a whole Java runtime — from package registries at run time and executed them, with no checksum standing between a registry compromise and your machine. The image closes that channel: nothing is fetched at run time. The language server version is pinned in this repository, every build dependency is pinned by SHA-256 in gradle/verification-metadata.xml, the base image is pinned by digest, and each release is an immutable image you can pin by tag or digest yourself. The container also bounds what the server can touch: it reads and writes only the project directory you mount.

The supply chain behind the build shrank with the move too. The npm package's lockfile pinned 288 packages, each one an independently owned npm account and a separate thing to trust; MCP's documented install idiom, npx -y <server>, additionally fetches the latest of all of that at launch, and that channel has already burned the MCP ecosystem — the September 2025 chalk/debug compromise reached transitive dependencies of the official MCP TypeScript SDK. This server's runtime classpath is about 35 JARs from a handful of organizations, JetBrains and Eclipse for the most part, and it changes only when a commit to this repository changes the pinned checksums.

Available Tools

10 tools
close_documentA

Close a file in the LSP server. Use this tool when you're done with a file to free up resources and reduce memory usage. It's good practice to close files that are no longer being actively analyzed, especially in long-running sessions or when working with large codebases.

Inputs: file_path (required).

get_code_actionsA

Get code actions for a specific range in a file. Use this tool to obtain available refactorings, quick fixes, and other code modifications that can be applied to a selected code range. Examples include adding imports, fixing errors, or implementing interfaces. Requires the file to be opened first.

Inputs: end_line (required), file_path (required), end_column (required), start_line (required), +2 more.

get_completionsA

Get completion suggestions at a specific location in a file. Use this tool to retrieve code completion options based on the current context, including variable names, function calls, object properties, and more. Helpful for code assistance and auto-completion at a particular location. Use this when determining which functions you have available in a given package, for example when changing libraries. Requires the file to be opened first.

Inputs: line (required), column (required), file_path (required), language_id (required).

get_diagnosticsA

Get diagnostic messages (errors, warnings) for files. Use this tool to identify problems in code files such as syntax errors, type mismatches, or other issues detected by the language server. When used without a file_path, returns diagnostics for all open files. Requires files to be opened first.

Inputs: file_path.

get_info_on_locationA

Get information on a specific location in a file via LSP hover. Use this tool to retrieve detailed type information, documentation, and other contextual details about symbols in your code. Particularly useful for understanding variable types, function signatures, and module documentation at a specific location in the code. Use this whenever you need to get a better idea on what a particular function is doing in that context. Requires the file to be opened first.

Inputs: line (required), column (required), file_path (required), language_id (required).

open_documentA

Open a file in the LSP server for analysis. Use this tool before performing operations like getting diagnostics, hover information, or completions for a file. The file remains open for continued analysis until explicitly closed

Inputs: file_path (required).

restart_lsp_serverA

Restart the LSP server process. Use this tool to reset the LSP server if it becomes unresponsive, has stale data, or when you need to apply configuration changes. Can optionally reinitialize with a new root directory. Useful for troubleshooting language server issues or when switching projects.

Inputs: root_dir.

save_documentA

Save a file in the LSP server for analysis. Use this tool before performing operations like getting diagnostics, hover information, or completions for a file. The file remains open for continued analysis until explicitly closed

Inputs: file_path (required).

set_log_levelA

Set the server logging level. Use this tool to control the verbosity of logs generated by the LSP MCP server. Available levels from least to most verbose: emergency, alert, critical, error, warning, notice, info, debug. Increasing verbosity can help troubleshoot issues but may generate large amounts of output.

Inputs: level (required).

start_lspA

Start the LSP server with a specified root directory. IMPORTANT: This tool must be called before using any other LSP functionality. The root directory should point to the project's base folder, which typically contains configuration files like tsconfig.json, package.json, or other language-specific project files. All file paths in other tool calls will be resolved relative to this root.

Inputs: root_dir (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.1/5.0

    Scored across 10 tools

    Disambiguation4/5

    Each tool addresses a distinct LSP operation, but save_document and open_document have identical descriptions, creating potential confusion about their respective roles. Hover, completions, code actions, diagnostics, and lifecycle tools are otherwise clearly separated.

    Naming Consistency4/5

    Most tools follow a verb_noun pattern with get_ for queries and open/save/close for document management. Minor inconsistencies exist: start_lsp lacks 'server' while restart_lsp_server includes it, and get_info_on_location is more verbose than sibling get_ tools.

    Tool Count5/5

    10 tools appropriately cover server lifecycle, document lifecycle, and core LSP queries without redundancy or bloat. This is well within the ideal 3-15 range for a focused MCP server.

    Completeness4/5

    The tool set covers the essential LSP workflow: start server, open documents, retrieve diagnostics, hover info, completions, and code actions. However, common LSP features like go-to-definition, find references, formatting, and rename are absent, leaving notable gaps for a full language server experience.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      Provides LLM clients with structured code intelligence through LSP servers, enabling queries for definitions, references, call hierarchies, and more.
      3
      Apache 2.0
    • A
      license
      Not graded
      quality
      D
      maintenance
      Provides AI agents with language-aware code analysis through the Language Server Protocol, enabling tasks like getting code insights and diagnostics.
      194
      MIT
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables AI coding agents and MCP clients to pull real-time language-server data from a running editor instance, including diagnostics, hover/definition/signature and type information, and symbol references with usage context. It also supports workspace-wide symbol renaming, opening files, and command execution, letting agents validate changes without waiting on slow tsc/eslint runs.
      -
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables AI coding agents to obtain real-time language-server diagnostics, type information, symbol definitions, references, and workspace-wide renames directly from open editor instances, avoiding slow external type and lint checks.
      -