ArchitectureA supervised JVM-class runtime — OLTP on seven engines, OLAP on three. AI-native, MCP-native, observable as plain SQL.Read the architecture
Está viendo la edición Perú. Está viendo la edición Colombia. You're viewing the Pakistan edition. Cambiar a la edición global →Cambiar a la edición global →Switch to the global edition →
Engineering

MCP server initialization payload restructured to fit client limits

The MCP server's InitializeHandler instructions are rewritten to stay within the 2 KB client limit: per-table write rules and the filesystem tree move from the initialization payload into individual tool descriptions, and fs_read now streams large files line-by-line.

MCP clients impose a size limit on the instructions block returned at session initialization. As the platform's MCP server added filesystem operations, database queries, and documentation lookup, the instructions block exceeded that limit on several popular clients. This release restructures the payload: only session-level context lives in the instructions field; per-tool constraints and filesystem-tree examples move into each tool's own description.

Instruction payload changes

  • Instructions capped at 2 KB. The initialization instructions now contain only the server identity, database scope, and top-level navigation model. Per-tool rules are absent from this block.
  • Per-table rules in tool descriptions. Write constraints for specific tables — which columns are settable, which require a prior read — now appear in the descriptions of the fs_write and fs_touch tools, reaching the agent exactly when it selects those tools.
  • Filesystem tree examples in tool descriptions. Navigation examples covering JavaScript modules, XDBL objects, and documentation pages moved from the instructions block into the relevant tool descriptions.

Tool-description fixes

  • Tool name alignment. Tool names referenced in instructions were audited against the actual tool names; mismatches caused agents to attempt calls the server rejected.
  • js:// resource reference expansion. The initialization handler now expands js:// URI references to resolved paths before embedding them in instruction text.
  • Large-file streaming for fs_read. File content is now streamed line-by-line rather than materialised in full, so large JavaScript modules and documentation pages do not exhaust the tool response budget.
  • apps/query read-only clarification. The tool description now states explicitly that this path is read-only, preventing agents from attempting mutations through it.

The net effect is a compact, authoritative session context at initialization and per-tool guidance at the point of use — matching how a developer reads API reference documentation rather than front-loading every constraint into the session preamble.

See the feature →

← All posts