Local hardware context for coding agents.
Axon is an open-source MCP server that reports CPU, memory, disk, thermal, GPU, process, and agent-runtime pressure before a local agent starts more work.
Engineering analysis, not marketing claims.
A public research log for real scenarios: same task with and without Axon, measured locally, with the tradeoffs shown instead of hidden.
With and without Axon: app build, test, helper fan-out, and output preservation.
The same app task was run in independent workspaces. Axon avoided extra tool helpers, reduced concurrency, preserved output, and made the tradeoff explicit.
Turning local runtime pressure into value metrics users can understand.
Risky tool spawns avoided, estimated credits saved, developer time protected, and stale local runtime count.
Why agentic workflows need host awareness before they run more tools.
OOMs, stuck sessions, large local logs, slow builds, and false debugging paths all point to a missing runtime signal.
A generic demo for agent IDEs, app builders, local runners, and GPU boxes.
How Axon sits beside any agent runtime and returns run, degrade, defer, cleanup, and safe_parallelism decisions.
The bottleneck moved below the prompt.
Modern coding agents can plan, edit, test, spawn tools, drive browsers, and call MCP servers. But they usually do it blind to the local machine they are running on.
Tool servers accumulate
Old MCP servers, stale shells, renderers, and browser controllers quietly compete with every new agent task.
Agents blame the code
When hardware pressure causes flaky tests or slow builds, the agent may debug false software problems instead of reducing local pressure.
Users feel the slowdown
The visible outcome is lag, OOMs, fan noise, slow IDEs, failed browser automation, and wasted tokens.
What Axon gives an agent
Example decision
$ axon query workload_advice
{
"recommendation": "defer",
"risk": "critical",
"safe_parallelism": 1,
"reasons": [
"Disk at 96% (Critical)",
"System impact is critical",
"Current anomaly is memory_pressure"
],
"suggested_actions": [
"defer new heavy work",
"cap parallelism at 1 instead of 4"
]
}Where it fits.
Axon is intentionally small enough to run next to local developer tooling and explicit enough for agents to turn host state into execution policy.
Preflight before tool fan-out
Check host pressure before browser runs, builds, test loops, MCP calls, and subagent work.
Policy for shared workstations
Return run, degrade, defer, cleanup, and safe_parallelism decisions without sending process data off-device.
Better failure diagnosis
Separate code failures from machine pressure, stale local processes, thermal throttling, and memory growth.
Open source by design.
Axon’s privacy boundary is part of the product contract. Community docs make that boundary reviewable before adoption.
Install once. Every local agent gets hardware context.
Axon is private by architecture: no telemetry, no analytics, no cloud calls. It speaks MCP over stdio and runs locally.
# Homebrew
brew install rudraptpsingh/tap/axon
# Configure agents
axon setup
axon setup cursor
axon setup vscode
# Try it directly
axon query agent_runtime_health
axon query workload_adviceRun the local check before heavy agent work.
Use Axon before builds, browser automation, Docker work, GPU jobs, subagents, and MCP-heavy sessions so the agent can cap, defer, or clean up with a concrete reason.