Hermes Agent
A self-hosted operations agent that runs around the clock.
What it is
Hermes Agent is Nous Research's open-source agent. Nix Inc runs its own copy on a 64 GB Apple M5 Pro Mac, with one gateway serving two profiles. The main profile runs on Claude Sonnet 5.5 with a 1M-token window, through a Claude Max subscription, and switches to GPT-5.6 Sol when Claude returns a rate-limit, overload or connection error. The second profile runs a model on the Mac itself, covered in its own build. Both take instructions over Telegram; scheduled jobs run under the main profile.
By the numbers
What it means for your team
It shows what an always-on agent needs to be dependable: a control channel on a phone, scheduled jobs, a backup model, written rules with approval gates, and a backup before config changes. Failures are logged with their cause so they can be fixed or watched.
Status
Nous labels the Claude plugin experimental. Sign-in failures stopped the agent on 2026-09-21 and 09-28; on 2026-10-08 the backup model covered one. A daily Maillard accuracy job was paused on 2026-10-08 after it kept failing on a missing model setting. Of the 356 skills installed, 198 came from the public skills hub and 127 are enabled. The behavior evals are rule checks against fixtures, not live model scoring.
Services
Hermes Agent v0.21.5 is installed from the upstream repository. Eleven macOS launchd jobs run the stack, including the gateway, the dashboard that ships with Hermes, Open WebUI, the local model servers and a small in-house web UI that keeps the API key on the server.
Model routing
The main profile uses Nous's official Claude subscription plugin, which drives the unmodified Claude Code CLI on a Claude Max login. Rate-limit, overload and connection errors fall back to GPT-5.6 Sol.
Control and schedule
Each profile has its own Telegram channel, and the main profile's bot answers only one allow-listed chat. Two scheduled jobs are enabled, the morning scan and the weekly deep dive, both set to run on Claude with GPT-5.6 Sol as the backup. During scheduled runs, any action that needs approval is denied automatically.
Guardrails
The agent's instruction file holds a written operating contract and writing rules derived from 336 sentences of its own replies. An in-house plugin adds 24 tools, including an evidence ledger, a decision log and a safety preflight that blocks secret file paths.
Tuning audit
From 2026-09-28 to 09-30, a four-part audit covered upstream changes, model capabilities, telemetry and config, including timing data from 672 logged Claude calls. Each approved fix was applied after a backup, and one test mismatch was reported on the plugin's public repository.