Context Management keeps MCP traffic from flooding the AI's context window. It does two independent jobs:

• Tool Responses Indexing — a tool response larger than the response threshold is indexed and replaced with a preview plus the exact calls that read or search the rest.
• MCP Catalog Indexing (catalog compression) — when the combined tool, resource, and prompt catalog is larger than the catalog threshold, server instructions and tool definitions are indexed and deferred behind search.

Both are backed by a per-session, in-memory SQLite FTS5 index (with optional hybrid vector search). The AI gets two tools to use it: IndexSearch and IndexRead (names are configurable).

Key benefits:

• Large tool responses no longer consume the context window, yet nothing is lost
• Small and medium responses (up to 16 KB by default) pass through untouched
• Large catalogs shrink to summaries; deferred tools are activated on demand
• Per-client enable/disable with global defaults

Catalog compression runs during MCP session initialization when MCP Catalog Indexing is on. It applies progressive phases, in order, until the estimated catalog size falls below the catalog threshold (default: 16,384 bytes).

IndexSearch queries the index. Pass every question in one call:

{
  "tool": "IndexSearch",
  "arguments": {
    "queries": ["create github issue", "file management"],
    "source": "mcp/"
  }
}

• Hits are ranked best-first, show line numbers, and come with up to ~3,000 characters of context each. A chunk already shown for an earlier query is referenced instead of repeated.
• limit sets hits per query (default 5, max 20).
• source keeps only sources whose label starts with it — mcp/ for the catalog, mcp/github for one server, jira__getIssue: for one tool's responses. A source that matches nothing returns an error listing the indexed labels.
• Deferred tools, resources, and prompts found by a search are activated, and the gateway notifies the client with notifications/tools/list_changed.

IndexRead returns a source by label as numbered lines — 200 lines by default, up to ~64 KB per call. When it stops early it ends with the offset to continue from:

[173 more lines — continue with offset="61"]

Lines longer than 2,000 characters are split into parts labelled N-M (part M of line N); limit counts whole lines.

When a tool response exceeds the response threshold (default: 16,384 bytes), the full output is indexed under a unique label — the tool name and a run number, e.g. filesystem__read_file:3 — and the response is replaced with:

[Response compressed — 34176 bytes, 233 lines, indexed as "filesystem__read_file:3"]

<first ~2 KB of the response>
[… preview ends at line 41 of 233]

## Contents
- [L1] Overview
- [L42] Endpoints

Read all of it: IndexRead(label="filesystem__read_file:3", limit=233)
Read a range:   IndexRead(label="filesystem__read_file:3", offset="<line>", limit=<lines>)
Search it:      IndexSearch(queries=["<terms>"], source="filesystem__read_file:3")

A response is left unchanged when the placeholder would not be smaller than the response itself. Run numbers are never reused while LocalRouter runs, so a label always names the same content.

Context Management is configured globally and can be overridden per client. Settings left at their defaults are not written to the config file, so improved defaults reach you automatically.