Agent profile

Hermes Agent

Self-improving open-source AI agent with persistent memory and learning loops that creates skills from experience; terminal interface plus messaging gateway (Telegram/Discord/Slack/WhatsApp/Signal).

Overview

Hermes AgentOverview
VendorNous Research
LicenseMIT
Latest version0.16.0v0.16.0 (2026-06-06)
GitHub stars195000APPROX snapshot; re-snapshot via GitHub API
LanguagePython~82% Python
InterfaceCLI, Desktop app, APINative app (mac/win/linux) + terminal + messaging gateway
PricingFree / OSSMIT; pay only model tokens

Agent capabilities

Hermes AgentAgent capabilities
TopologySingle agentSelf-improving; skill creation
PlanningReflection loopBuilt-in learning loops
Tool callingMCPMCP + native tools
ExtensibilityMcp servers, Custom tools, HooksSkills system
MemoryPersistent db, SummaryFTS5 session search + LLM summarization; Honcho user modeling
Code executionNative
Browser/computer useVerify browser/computer-use support
MCPNative
MultimodalTextVerify vision/audio

Models

Hermes AgentModels
Model-agnosticYes
ProvidersOpenAI, OpenRouter, OtherNous Portal, OpenRouter (200+), NovitaAI, NVIDIA NIM, z.ai/GLM, Kimi/Moonshot, MiniMax, HuggingFace, or own endpoint
Local modelsVia proxyAny OpenAI-compatible endpoint
BYO keyYes

Self-hosting

Hermes AgentSelf-hosting
Self-host difficultyEasyFIRST-PARTY: official install.sh one-liner (the same script our production preset runs), then provider API key config; see /learn/self-host-hermes-agent
DeployNativePython; runs on any small VPS per public guides; confirm first-party
ScalingSingle processVerify
Config complexityModeratemodel config is a YAML mapping (provider/default/base_url/api_mode), not a simple string; wiring a self-hosted OpenAI-compatible endpoint requires knowing that OPENAI_BASE_URL is honored only for the openai-api provider (not the generic "custom" provider) — a non-obvious caveat only confirmed via a direct docs full-text-dump grep after an earlier AI-drafted citation about this mechanism turned out to be fabricated
Min vCPU (measured)1 vCPUFIRST-PARTY: measured on a DigitalOcean s-1vcpu-1gb (1 vCPU / 1 GB, Ubuntu 22.04), harness run 20260826-195645, 2026-08-26. Installed, verified and answered a live completion on a single vCPU. The install is the slow part, not the running: 300s on one core versus 240s on two.
Min RAM (measured)1 GBFIRST-PARTY: measured on a DigitalOcean s-1vcpu-1gb (1 vCPU / 1 GB, Ubuntu 22.04), harness run 20260826-195645, 2026-08-26. Proven floor: all tiers passed on a 1 GB box including a live model completion, with 500 MB still available at idle. Below 1 GB is untested.
Idle RAM (measured)120 MBFIRST-PARTY: measured on a DigitalOcean s-1vcpu-1gb (1 vCPU / 1 GB, Ubuntu 22.04), harness run 20260826-195645, 2026-08-26. Resident set of the hermes gateway process (a Python venv) 30s after a completion; whole-box use was 460 MB. Barely moves under load — the working-set sample was 122 MB — which is the clearest contrast with openclaw in this data.
Infra costVery lowFIRST-PARTY: measured on a DigitalOcean s-1vcpu-1gb (1 vCPU / 1 GB, Ubuntu 22.04), harness run 20260826-195645, 2026-08-26. The lightest resident agent measured: ~120 MB, flat under load, leaving half of a 1 GB box free. Its cost is install time, not runtime.
One-click on AgentOceanYesFIRST-PARTY: production install preset in src/lib/products.ts (id: hermes) — runs the official install.sh at checkout

Project health

Hermes AgentProject health
MaturityBetaPre-1.0 (v0.16)
MaintenanceActive
Release cadenceMonthlyEstimate from version cadence
DocsGoodEstimate
CommunityLargeEstimate from star signal

A dash (—) marks a value we have not yet verified first-hand. Record last verified 2026-06-17.

Run Hermes Agent on your own VPS — one click

Hermes Agent is a one-click install on any AgentOcean VPS: pick a plan, tick the preset, and it's live in minutes.

Hermes Agent head-to-head