Update after more testing: the tools are not gone. The antigravity agent (gemini-3.1-pro-high) can still call AgentsRoom-MCP tools after a restart, but only through a lazily loaded call_mcp_tool bridge (ServerName: "AgentsRoom-MCP"). The agent had to load that bridge first. Until it did, it reported "missing MCP tool bindings" and listed the tools from memory without being able to call them. So the bug is that a restarted antigravity agent doesn't load the MCP bridge on its own, and says nothing useful about it. It is not that the tools are missing. Impact is lower than I first reported, but unattended trigger runs could still fail on this. A scheduled run has nobody to tell the agent to load the bridge. Suggested fix: load the bridge at session start, or tell the agent in its startup context that MCP tools are reached through call_mcp_tool.
Antigravity (Gemini) agent loses the MCP tool bridge after a restart and cannot call AgentsRoom tools
From PaulterchiefDone 2 days agovia agent-mcp
An agent on the antigravity provider (gemini-3.1-pro-high), used as a code reviewer, cannot call AgentsRoom-MCP tools after its session is restarted. Other agents in the same project on the claude and codex providers call the same tools without trouble.
Observed:
- The agent was asked to review a backlog ticket (it needed to call backlog_review). Its last status said it was "missing MCP tool bindings after restart".
- The user restarted the agent and asked it to list its backlog_* tools.
- After the restart it listed all the tools correctly by name and description. It also said "the call_mcp_tool bridge is currently not loaded in my active tool declarations" and told the user to reopen the project or refresh MCP settings.
Expected: after a restart the agent can invoke the AgentsRoom-MCP tools (for example backlog_review). Actual: the agent can describe the tools but cannot call them. The ticket stays in review because only that agent's role may approve it.
Steps: run an antigravity agent in a project with AgentsRoom-MCP. Restart its session. Ask it to call backlog_get or backlog_review.
Not verified: whether this happens on every restart, or only after a model switch. The agent's own note mentions "server restart and model switch".
Impact: a second antigravity agent is planned as a blog compliance reviewer and needs the same tools.
AgentsRoom diagnostics
- App version: 1.222.0
- OS: darwin 25.6.0 (arm64)
- Agent CLI: claude
- App language: system
Expected behavior
After a restart the antigravity agent can invoke AgentsRoom-MCP tools such as backlog_review.
Actual behavior
Agent lists the tools but reports the call_mcp_tool bridge is not loaded in its active tool declarations, so it cannot call them.
Steps to reproduce
Run an antigravity agent (gemini-3.1-pro-high) in a project with AgentsRoom-MCP. Restart its session. Ask it to call backlog_get or backlog_review.
Replies
Fixed, available in the next version. Cause: Antigravity reaches every MCP tool through a single bridge tool (call_mcp_tool) that it can load on demand, but nothing in the agent's startup instructions said so. After a restart it saw the tool names without seeing a way to call them, and reported the bindings as missing instead of loading the bridge itself. Antigravity agents are now told at startup that AgentsRoom tools go through call_mcp_tool (server AgentsRoom-MCP, short tool name), and that they must load that bridge on their own, which also covers unattended trigger runs. If it still happens after this update, please reopen the ticket with as much detail as possible (a screenshot or video helps a lot).
Thanks: the bug is fixed and the fix ships in the next release. If it still happens after that, tell us and we will gladly take another look.
Sent automatically with your message
Your browser sent this along so the team can reproduce what you saw. Nothing here was typed by you.
- Browser
- AgentsRoom-MCP/1.222.0 (darwin arm64)