Context Mode

What the context-mode MCP server does

The context-mode MCP server is designed for coding sessions where tool output consumes too much of the model's context. Rather than placing every raw result in the conversation, it provides tools that execute code, process data, index content, and return selected results. The README describes reductions such as turning hundreds of kilobytes of raw data into a few kilobytes of output, though the actual savings depend on the task and the way tools are used.

It also preserves session state. File edits, Git operations, tasks, errors, and user decisions are recorded in a SQLite database. When a conversation is compacted, the stored events are not all reinserted into the prompt. Relevant information can instead be located through full-text search.

The system does not enforce a particular style for the model's final response. Its routing guidance concerns where data should be processed, not whether the agent should be brief, verbose, or formatted in a specific way.

How it works

The context-mode MCP server exposes sandbox tools for running JavaScript, running a script file, processing batches, indexing local content, searching indexed material, and fetching and indexing content. A coding agent can use these operations to calculate results close to the data instead of reading many files into the conversation. For example, a script can count lines across a directory and print only the totals.

Indexed session information uses SQLite with FTS5. Search uses BM25 ranking to retrieve relevant events rather than restoring the entire prior session. A continued session can therefore recover selected context after compaction. If continuation is not requested, the README says previous session data is deleted, providing a clean session instead.

Routing is enforced through hooks on supported hook-capable platforms. The project describes this routing and code-oriented workflow as applying across its supported clients, as well as through an OpenClaw gateway integration.

Setup and configuration

Claude Code can install the project through its plugin marketplace. The documented flow is to add the mksglu/context-mode marketplace, install the context-mode@context-mode plugin, and restart Claude Code or reload plugins. Claude Code version 1.0.33 or newer is required.

The Claude Code plugin registers hooks for session start, tool use, prompts, compaction, and stopping. Its session-start hook supplies routing instructions at runtime rather than writing a routing file into the project. The plugin also registers the MCP tools and provides a diagnostic command that checks runtimes, hooks, FTS5, and plugin registration.

On other platforms, the README states that non-hook clients require a one-time routing-file copy, while hook-capable clients receive automatic routing enforcement. The supplied material does not specify the exact file contents or installation commands for those platforms.

Tools and capabilities

The documented Claude Code installation includes 11 MCP tools:

• ctx_batch_execute, ctx_execute, and ctx_execute_file for sandboxed computation.
• ctx_index, ctx_search, and ctx_fetch_and_index for building and querying the local knowledge base.
• ctx_stats for context-use and savings information.
• ctx_doctor for diagnostics.
• ctx_upgrade for updating and repairing the installation.
• ctx_purge for permanently removing indexed content.
• ctx_insight for opening the hosted Insight dashboard in a browser.

The same functions can be reached through Claude Code slash commands. The README says other platforms can invoke the utility functions by typing commands such as ctx stats, ctx doctor, ctx index, or ctx search in chat.

Limitations and notes

This tool changes how data is routed and processed; it does not guarantee a fixed reduction for every workload. Agents still need to use the sandbox and indexing tools appropriately. The project also does not replace a model's response-style instructions. Indexed content can be permanently removed with ctx_purge, and sessions that are not continued lose their previous stored data according to the documented behavior.