Local Agent Telemetry Report 2026: what sixteen coding agents send, and how to stop it

A local model does not make the agent around it local. This second-edition report from Stridenalysis, the research desk of StrideNote, tests sixteen coding and task agents: OpenJarvis, OpenCode, Pi, Hermes Agent, DeepSeek Harness, Cline, OpenClaw, Codex CLI, Aider, Kilo Code, Goose, OpenHands, Grok Build, ZeroClaw, smolagents and Google ADK.

Each agent ran on one test machine (Apple M4 Pro, 48 GB, macOS) against the same local model, Gemma 4 31B QAT served by LM Studio, with no account signed in, on the same small task. Every request was recorded and answered on the machine by a capture proxy in hold mode, so no test data reached the vendors. Every finding carries an evidence grade (A capture, B sockets, C code).

The task prompt never left the machine for any of the sixteen. Five agents report usage without asking. DeepSeek Harness 0.1.5 uploads the whole session when an answer is rated. The Cline command line tags every event with a hash of the machine's hardware UUID and has no switch for it. Hermes Agent refreshes Claude Code's stored sign-in at launch. Seven agents can be made fully quiet with settings their makers document. The report records seven false leads, five things that went wrong in our own testing, the switches for each agent, its limitations, and the tests planned for the next edition.

Files: the report (PDF, 48 pages) and the dataset behind every chart (agent-telemetry-2026.csv, one row per finding).