Stridess/local-agent-telemetry-2026

Local Agent Telemetry 2026

What sixteen AI coding agents send over the network when they run with a local model, measured by StrideNote on one Mac between 8 July and 22 September 2026. One row per agent per destination: 41 rows, 17 tools (one installer counted separately), 36 graded A, 4 graded B, 1 graded C.

This is the dataset behind the Local Agent Telemetry Report 2026 (edition 2, 48 pages, DOI 10.5281/zenodo.22969237). The report explains the method in full; this card gives what you need to read the table.

What was measured

Each agent was installed in an isolated profile on a 48 GB M4 Pro Mac, pointed at a local model server (LM Studio or Ollama), and run through the same short coding task. During launch, idle and the task we recorded every outside socket (sampled at 0.25 s), and ran the same steps again behind a capture proxy that held every request so nothing reached a vendor. A marker string placed in the prompt was searched for in every captured request to answer the question that matters most: does the work leave the machine?

Columns

column	meaning
tool	the agent, as named by its maker
version	the version tested
date	the test date
evidence_grade	A: captured at a listener or proxy we controlled. B: captured sockets plus source reading. C: source reading only.
method	how the row was obtained
destination	the host and path the agent contacted
trigger	when it happens: install, launch, during use, on a user action
what_is_sent	the payload, from the capture
content_of_work_sent	whether any of the prompt or files left the machine
on_by_default	yes or no
off_switch	the setting, flag or environment variable that stops it, if one exists
off_switch_documented	whether the maker's documentation names that switch
off_switch_verified	whether we ran the switch and saw the traffic stop
source_url	the StrideNote piece or report page the row comes from

Headline findings

• The prompt stayed on the machine for all sixteen agents.
• Five agents report usage without asking first.
• One agent's CLI ignores DO_NOT_TRACK and has no telemetry switch of its own.
• One agent derives its device identifier from the Mac's hardware UUID, so it survives reinstalls.
• One agent uploads the whole session when the user rates a reply.

Every figure above is in the table with its capture named.

Limits

One machine, one operating system, one short task per agent, on the dates given. Agents change quickly; a row describes the version tested, not the tool today. Socket sampling at 0.25 s can miss a shorter connection; the proxied runs cover anything that honours HTTPS_PROXY. Grade C rows come from reading source and were not captured.

Updates since the report

• 19 October 2026: Cline's documentation, revised on 7 October, now names an off switch for its command line, telemetryOptOut in its global settings file. The row here describes CLI 3.0.64, which had none; the CLI is now at 3.0.73. Not retested.
• 19 October 2026: OpenJarvis merged a change on 14 October that makes its analytics opt-in. The row here describes desktop-v1.0.2, still the newest desktop release on that date. Not retested.

Licence and citation

CC BY 4.0. Cite as: StrideNote (2026). Local Agent Telemetry Report 2026, edition 2. https://doi.org/10.5281/zenodo.22969237

Corrections: hello at stridenote.net, or the report page above.