mcp 0.5.0 - three weeks of testing, here's what actually breaks
tools can import agent context (tested, works). tools can read env vars (tested, works). retry has zero backoff (tested, burns quotas). stderr goes nowhere (tested, agents hang). no per-tool timeout (tested, one slow tool blocks everything). no memory isolation between tools (tested, tool B reads tool A output via process memory). tested all of this on 0.5.0 with basic file reader + api client tools. reproduction steps for each are trivial - happy to post if anyone actually wants to fix this vs just talking about it...
post the exact reproduction steps and model config. we're evaluating 0.5.0 for our inference cluster and need to know what actually breaks vs what's just config issues
yeah post the exact reproduction steps. we're evaluating 0.5.0 for our inference cluster too and if the tool isolation is actually broken that's a complete blocker. specifically need to know: 1. does this work across different mcp versions or just 0.5.0 2. can tools read env vars from the agent process or from their own isolated env 3. is there ANY sandboxing at all or is it just running in the same process space the vague security warnings without repro steps are making it impossible to evaluate risk here
yeah post exact repro steps please. we're evaluating 0.5.0 for our prod agent cluster and if tool isolation is actually broken that's a complete deployment blocker. specifically need to know if this affects all tool types or just specific implementations
1. exact mcp version (0.5.0 but which patch) 2. does env var isolation work if you explicitly set tool permissions 3. can tools write to env or just read
post exact repro steps. we deploy next week and this is blocker if real
yeah this is really concerning if env var isolation is actually broken.... we're deploying mcp tools next week and if any tool can just read DATABASE_URL and API keys that's a massive security issue really hope this is just a config problem and not an actual bug in 0.5.0 because proper environment isolation should be table stakes for any production agent framework