What happened
The Earshot GitHub release page lists version 0.4.1 as “community fixes,” released on November 6. Earshot is an audio tool intended for coding agents. Its earlier release notes describe three ways to use it: a command-line interface, a Python API, and an MCP server. The 0.4.1 notes thank four contributors and frame the update as a fix release after version 0.4.0.
Installation and updates now have one command each. The release page gives an installation command using uv tool install with Earshot's MCP extra from the project repository, followed by uv tool upgrade earshot for updates. Under uv, PyTorch comes from the CPU-only index. This matters for the project’s stated CPU-oriented setup: users do not need to infer a separate package source from the release notes. The page does not announce a price, hosted service tier, or changed access terms.
Several changes address cases where an automation could receive a misleading result. A NaN or Inf sample now produces a clear error. Before this change, the release notes say that most commands crashed, while match, rank, and realness could return wrong answers with exit code 0. In an agent workflow, an exit code of 0 can be treated as successful completion, so that prior behavior could allow a downstream step to proceed on an invalid result. The release also says media glob patterns now match files only.
The video.cuts behavior is tightened as well. Setting cache=False now actually rescans. A failed scene scan is reported as an error instead of “no cuts,” and a scan is cached only when it ran. These are modest changes, but they define more clearly whether a result is a real analysis, a reused result, or a failed operation. The release says that on valid input every command produces the same results as version 0.4.0, so the stated aim is reliability rather than changed output quality.
Resource handling is the most concrete operational change. BLAS threads now follow EARSHOT_THREADS in both the CLI and MCP server. On an idle MCP server running on a machine with 12 logical CPUs, the project reports reserved memory falling from 411 MB to 154 MB and the thread count falling from 21 to 13. The notes also report that Earshot recognizes real sound effects by their own titles in 168 of 195 cases. They caution that job-like wording, such as “something appearing,” is not heard as a sound; their practical recommendation is to describe a sound two or three ways and keep the best result.
What the archive adds
New Runtime's earlier coverage on October 19 described Earshot as a CPU-only tool that gives coding agents operations for working with sound. This is a new fact on that known story: version 0.4.1 concentrates on installation, error handling, caching, and idle MCP resource use rather than expanding the audio feature set.
Trend signal
This is a small instance of the agent protocols become the interoperability layer pattern. Earshot makes its audio operations available through an MCP server alongside a CLI and Python API, and this release improves the operational behavior of that server. One project’s maintenance release does not establish that MCP has become the durable integration boundary across agent tooling, but it shows that runtime costs and failure semantics matter once a tool is exposed that way.
Why it matters
A tool can be callable by an agent yet still be expensive to operate in a local workflow. Installation ambiguity slows onboarding, while idle memory and thread use can become material when a developer machine or CI runner holds several agent services at once. Earshot’s reported change—from 411 MB reserved and 21 threads to 154 MB and 13 threads on one 12-logical-CPU machine—gives existing users a specific reason to rerun their baseline after updating.
The error-handling changes may be more consequential than their patch-release label suggests. An explicit error gives an orchestration layer a chance to stop, retry, or ask for input. A wrong answer with a successful exit code is more likely to pass unnoticed into later work. For builders using Earshot as one component in a chain, the update improves the distinction between a completed analysis and an invalid or failed one. It does not remove the need to verify audio judgments before using them in production decisions.
Limits and uncertainty
The memory and thread comparison covers one idle MCP server on one machine with 12 logical CPUs. The release does not measure active audio processing, concurrent server instances, startup time, or behavior on other hardware. The project-reported 168-of-195 recognition figure is not an independently checked benchmark and does not show performance on arbitrary prompts, recordings, or production audio libraries. Earshot says valid-input outputs match 0.4.0, but does not quantify throughput or latency under load. No new pricing, access expansion, or audio-generation capability is announced.
Skill to build now
If Earshot is already part of a coding-agent workflow, set an explicit EARSHOT_THREADS value and record idle memory, thread count, and a small set of known audio tasks before and after upgrading. Include invalid audio containing NaN or Inf in that smoke test, plus a video.cuts run with cache=False. This turns the release’s claims into an environment-specific acceptance check rather than an assumed improvement.
Reader takeaway
Earshot 0.4.1 is a maintenance release for existing users, not a reason to adopt a new audio stack on its own. It reduces reported idle MCP overhead, makes installation more direct, and replaces several ambiguous failures with clearer behavior. Test the resource figures and audio results in your own agent environment before treating them as general performance guarantees.