FlowtandBriefAI · Software · ResearchMeet Flowtand

Reachy Mini adds remote MCP tools hosted in Hugging Face Spaces

Announcement

Hugging Face

Browse Agents topic page →

A Hugging Face blog post describes how the Reachy Mini conversation app can now call tools hosted in public Gradio Spaces over MCP. One command validates a Space, probes its MCP endpoint, discovers its tools and appends them to a profile's tools.txt. The post frames this as a way to add capabilities such as weather and web search without shipping Python into the local app, and notes limits around private Spaces, non-Gradio Spaces and guaranteed parallel tool calls.

Remote tools via MCP Spaces

The Reachy Mini conversation app now supports a third tool category alongside built-in robot tools and custom local tools: tools hosted in public Hugging Face Spaces and called over MCP. According to the post, adding one takes a single command that validates the Space on the Hub, probes its MCP endpoint, discovers its tools and, by default, appends the resulting tool IDs to the active profile's tools.txt. The tool itself keeps running in the Space, so no code is downloaded locally. tools.txt remains the gatekeeper: a remote tool is only active if its namespaced ID (for example a Space slug with slashes collapsed to underscores, followed by a double underscore and the tool name) is listed there. The post names two canary Spaces for search and weather, and notes that duplicate tool names across the merged set cause the app to fail fast.

Scenario: a demo-booth assistant

A developer building a booth demo could try a profile that keeps expressive body tools (emotions, head movement) and adds a remote search Space for current-information questions, following the canary_web_search pattern described in the post. A team presenting this to visitors could phrase the demonstration as answering a time-sensitive question aloud rather than as a product claim. Because the post says prompts decide how tools are used, the demo script should include instructions for when to call search and how to answer briefly in spoken style. This remains conditional: the post describes the canary profiles and two test Spaces, not a validated public deployment.

Trial steps and success checks

A team could start with one remote tool rather than two, install it with --install-only first to confirm discovery succeeds, then enable it in a dedicated profile. Baseline: the same brief of questions answered by the model without the remote tool. Trial: the same questions with the tool enabled and prompt instructions adapted from the canary examples. Success would mean the tool is invoked when current information is needed and the spoken answer stays short; failure checks could include install errors when discovery fails, and the app refusing to start on duplicate tool names. Because the post says parallel calls are not guaranteed, the team should inspect whether mixed weather-and-news questions serialise calls, and keep a human reviewing answers before showing them to an audience.

Trade-offs and when to skip

The trade-off is reach versus control: remote tools are easy to share and update independently, but the post restricts them to public, MCP-compatible Gradio Spaces. Private or authenticated Spaces, non-Gradio Spaces and arbitrary raw MCP URLs are listed as unsupported, so anything sensitive or non-standard is out of scope. Latency matters too: the post warns that vague prompts can serialise calls and add delay, and that deterministic orchestration belongs in code rather than in prompts. A team should not adopt this path if it needs authenticated data sources, guarantees about call ordering, or a fully local toolchain; in those cases local tools or custom external_tools remain the described options.

Context and limitations

  • Based on a short source excerpt; the full article was not analyzed.
  • The post lists unsupported cases: private or authenticated Spaces, non-Gradio Spaces, arbitrary raw MCP URLs and guaranteed parallel tool orchestration.
  • Prompt instructions can encourage parallel tool calls but cannot guarantee them; the post suggests moving deterministic logic into code.
  • Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.