Hermes Agent
Self-hostable AI agent with persistent memory, agent-written skills and a gateway to 20+ chat platforms.
AI
Hermes Agent is an autonomous AI agent from Nous Research that you install and run on your own machine or server (or, optionally, on the vendor's Hermes Cloud). It keeps persistent memory across sessions, writes and improves its own skills after complex tasks, exposes one messaging gateway to 20+ chat platforms, runs natural-language cron jobs unattended, delegates work to isolated subagents and connects to MCP servers for extra tools. It ships as a CLI, a desktop app and an official Docker image, and works with any OpenAI-compatible provider — OpenRouter, OpenAI, Anthropic, Google, a local Ollama/vLLM server, or the vendor's Nous Portal subscription, which is the recommended one-login setup and starts at 20 USD per month. The software itself is free and MIT-licensed: you pay your chosen model provider for inference, or nothing at all with a local model.
Key features
- Self-HostedFree self-hosted
- LicenceOpen source · MITDeploymentSelf-hosted
Limits, notes & source
Self-Hosted
- Limits
- Full agent feature set: persistent memory, agent-created skills, messaging gateway, cron scheduling, subagent delegation, MCP servers and all terminal backends. You supply the machine and pay your chosen LLM provider for inference; the recommended one-login setup (Nous Portal) is a paid subscription from 20 USD per month, and a local model endpoint needs a context window of at least 64,000 tokens.
- Notes
- The FAQ states the agent itself is free and open source under the MIT licence and that you pay only the LLM provider; the vendor's landing-page FAQ adds that cloud hosting is optional. Licence confirmed in the repository LICENSE file (MIT, Copyright (c) 2025 Nous Research). No user limit is published. Nous Portal is deliberately not recorded as an edition: it is a subscription for model and tool credits used with a self-hosted agent, not a way to obtain the software.
- Hermes Cloudper instance
- LicenceLicence unverifiedDeploymentManaged cloud available
Limits, notes & source
Hermes Cloud
- Limits
- Two instance sizes, priced per day and excluding inference and tool usage: Medium at 0.56 USD/day running and 0.03 USD/day stopped (10 concurrent sessions, 2 GB RAM, 4 vCPU); Large at 1.09 USD/day running and 0.03 USD/day stopped (20 concurrent sessions, 4 GB RAM, 8 vCPU). Inference and tool usage are billed separately on top, and a 2 USD minimum in credits or an active subscription is required to deploy.
- Notes
- Billed daily in arrears from Nous credit, metered to the second of actual uptime. No monthly or yearly figure is published and there are two published sizes, so no single amount is recorded — the per-day rates are carried here instead. Hosting is a service, so no open-source licence applies to this edition.
Pros & cons
Strengths
Trade-offs
Show 1 more
- A local model endpoint needs a context window of at least 64,000 tokens. Source
Hermes Agent replaces
Products this app is built to take over from, with the vendor's own evidence.
Moving your content across: Documented per product: what you do on the side you are leaving, what this app does with the result, and what does not survive the trip. Every step links to the page that documents it.
- Importhermes claw migrate previews every item, then imports persona (SOUL.md), memory (MEMORY.md/USER.md), skills, and model and provider config from ~/.openclaw; legacy ~/.clawdbot is detected too.Source
- CaveatSecrets are never imported silently — even --preset full needs --migrate-secrets — and a pre-migration backup zip of ~/.hermes is written unless you pass --no-backup.Source
- CaveatWorkspace instructions only land where you point them: AGENTS.md needs --workspace-target, and skill name clashes follow --skill-conflict (skip by default).Source
- Claude CodeA one-command importer reads an existing Claude Code install, mapping instructions, allowlists, MCP servers, skills and memories.Source
- Importhermes import-agent claude-code (auto-detects ~/.claude) turns CLAUDE.md into memory entries, Claude's Bash permission rules into command_allowlist and approvals.deny, and copies MCP servers and skills.Source
- CaveatCredentials are never read, MCP environment variables and headers with secret-looking names are stripped, and slash commands under commands/*.md are skipped with a note — rewrite them as skills.Source
- OpenAI Codex CLIThe same documented importer reads a Codex CLI setup, including AGENTS.md, MCP servers, memories and skills.Source
- Importhermes import-agent codex reads ~/.codex: AGENTS.md and memories/*.md become memory entries, config.toml's [mcp_servers.*] becomes mcp_servers, and skills go to ~/.hermes/skills/codex-imports/.Source
- Caveat~/.codex/auth.json is never read, and secret-looking MCP environment variables are stripped and listed in the report so provider credentials get re-added deliberately.Source
Similar apps
Other entries from this catalogue in the same category, with a free self-hosted edition first and then the higher star count. Only software we publish appears here — each card opens the entry, where its price, licence and check date are recorded.
