Skip to content

Known Limitations

Current limitations of SynContext and the MCP ecosystem. These are documented for transparency and will be addressed as the ecosystem matures.

MCP Client Tool Caching

Status: Ecosystem limitation (not a SynContext bug) Affects: All MCP clients (ChatGPT, Claude Desktop, Claude Code, Cursor, etc.) Impact: When SynContext adds new MCP tools in an update, previously connected clients don't see them until reconnected.

Background

The MCP protocol (Model Context Protocol) defines a notification mechanism called notifications/tools/list_changed that allows servers to inform connected clients when the tool list changes. SynContext's server framework supports this capability.

However, as of April 2026, most MCP clients do not handle this notification:

Client Handles list_changed? Workaround
ChatGPT No — freezes tool list at connection time Disconnect and reconnect in Settings > Apps
Claude Desktop No — reads tools once at startup Restart the application
Claude Code No — calls listTools only on initial connection Remove and re-add the MCP server
Claude.ai (web) Yes — refreshes on reconnection Start a new conversation
VS Code (Copilot) Partial — feature request open Restart VS Code
Cursor Unknown Restart the application
Gemini CLI No — feature request open Restart the CLI

What SynContext does

SynContext declares the listChanged capability in its server configuration. When MCP clients add support for notifications/tools/list_changed, tool updates will propagate automatically without requiring reconnection.

What you can do

When you see a SynContext release note mentioning new MCP tools, reconnect your AI clients:

  • ChatGPT: Settings > Apps > Disconnect SynContext > Reconnect
  • Claude Desktop: Quit and reopen the application
  • Claude Code: claude mcp remove syncontext then re-add
  • Other clients: Restart the application

See Troubleshooting for detailed steps.


GitHub Integration Rate Limits

Status: GitHub API limitation Affects: hub_github_search tool Impact: GitHub's Search API has a separate rate limit of 10 requests per minute, independent of the general 5,000 requests/hour limit.

If you hit the search rate limit, wait 60 seconds and try again. Other GitHub tools (tree, read, commits) use the general rate limit and are unlikely to be affected.


ChatGPT Confirmation Dialogs

Status: ChatGPT design decision Affects: All MCP tool calls from ChatGPT Impact: ChatGPT requires user confirmation for every external tool call. This is by design and cannot be disabled.

Unlike Claude, which can chain multiple SynContext tool calls automatically, ChatGPT will pause and ask for confirmation on each one. This makes multi-step workflows slower but is a security feature of the ChatGPT platform.


Legacy API Key (__legacy__)

Status: Deprecated (will be removed in a future version) Affects: Users of the static HUB_API_KEY environment variable Impact: The legacy API key authenticates as tenant __legacy__ which does not exist as a real user in PostgreSQL. All MCP tools will fail with "internal error."

Solution: Register a user account at syncontext.dev/app and use the API key from your account settings.