Hermes Agent has always explained that if the tool is not called, it depends on three things: whether the current model supports tool calls, whether the platform toolset is enabled, and whether the task is blocked by permissions or security policies. This question has been a frequent topic in community discussions recently, and many times it is not that Hermes is broken, but that the model or configuration does not give it the conditions to "do it".
Step 1: Confirm whether the tool is open
Run hermes tools to see the tools enabled by your current platform. Portals such as CLI, Telegram, Open WebUI, Discord, etc. may have different toolsets, and just because a platform can chat doesn't mean you can execute files, terminals, browsers, or MCP tools.
Step 2: Confirm whether the model can be adjusted or not
Some open-source models or compatible interfaces will say they support the OpenAI format, but function calls, tool schemas, and streaming returns are incomplete. The performance is that the model knows you want it to do things, but only explains it and does not really call the tool. At this time, calling a more stable model test with another tool is the fastest way to locate it.
Step 3: Look at permissions and approvals
- Dangerous orders may require your approval, and the messaging platform confirms that the approval button or receipt can be delivered.
- Backends like Docker, SSH, Modal, etc. may not have the target directory mounted, causing the tool to not see the files.
- External tools such as browsers, web searches, and MCPs also depend on whether the corresponding API keys and dependencies are configured.
Do not open the --yolo directly during troubleshooting. First, let Hermes perform a low-risk action, such as "list current directory files", confirm that the tool link is open, and then gradually release permissions. This will not only identify the problem, but also avoid removing the security boundary for configuration repairs.
