Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.mcp-assistant.in/llms.txt

Use this file to discover all available pages before exploring further.

v2.0.0 includes breaking changes. See the migration guide before upgrading.

New features

Neon storage backend

A new Neon storage backend persists MCP sessions on Neon’s serverless Postgres. It ships with optional Row-Level Security configuration and a pg_cron cleanup job for expired sessions. See the storage backends overview for how it compares to Redis, Supabase, SQLite, and the file system.

Code Mode

Code Mode is a new execution mode that lets the LLM write and run small TypeScript snippets in an isolated sandbox. The model calls your MCP tools as plain async functions (for example, github.get_issue(...)) instead of issuing one tool call at a time. Code Mode auto-generates TypeScript interfaces for connected servers and ships four built-in tools for searching, listing, inspecting, and chaining tool calls. The sandbox runtime is an optional dependency.

Tool Router as a standalone package

The Tool Router is now available as its own installable package, so you can drop it into any agent or framework without pulling in the rest of mcp-ts. The default meta-tools behave exactly as before.

Updates

Standardized session API

v2.0.0 aligns the session and client APIs around consistent naming. Highlights:
  • The user identifier passed to clients, storage backends, and hooks is now userId (previously identity).
  • Session management has moved from storage to a dedicated sessions API with create, list, get, and update methods.
  • The Streamable HTTP transport type is now spelled streamable-http consistently across config, client, and server options.
  • Session data accessors on storage backends have been renamed for clarity β€” see the storage reference.
Database-backed storage (Supabase, Neon, SQLite) requires running the provided migrations before deploying. Full upgrade steps are in the v2.0 migration guide. mcp_search_tools now indexes words from nested argument schemas β€” object properties, array item types, and enum values β€” not just the top-level tool name and description. The LLM can find a tool by terms that appear only inside its input schema, improving recall on large MCP servers without changing the response shape.

Tool approvals and annotations

The tool approval flow now delivers a more consistent approval payload to apps that use the React or Vue hooks. We also removed annotation fields from tool definitions to simplify the tool types.

Optional peer dependencies

json-schema-to-zod and isolated-vm are now optional. Install them only if you generate Zod schemas from MCP tool schemas or run Code Mode, respectively, which keeps the default install slimmer.

Bug fixes

  • SSE header construction β€” The SSE client now trims header names and values and skips empty entries, preventing malformed optional headers from leaking into connection setup.
  • Tool search recall β€” Fixed cases where nested schema fields prevented matches on otherwise valid search terms.