Jonathon's AI Wiki

Nous Research Hermes Agent

5 min read

Source: Nous Research — Hermes Agent site + docs + repo Repo: github.com/nousresearch/hermes-agent (MIT, 97K+ stars, Python, default branch main) Homepage: hermes-agent.nousresearch.com Version (2026-04-17): v0.10.0

Nous Research’s autonomous agent framework. Positioned against two contrasting categories: coding copilots (“tethered to an IDE”) and chatbot wrappers (“around a single API”). Hermes is neither — it’s a long-running agent that installs on a server you control, curates its own memory, generates its own skills, and gets more capable the longer it runs. MIT-licensed; 60-second install on Linux, macOS, or WSL2.

Key Takeaways

  • Lives on your server, not in an IDE or SaaS dashboard. The architectural bet is persistence: the agent that remembers how it solved a problem last Tuesday outperforms the one that starts fresh every session.
  • Closed learning loop. Agent-curated memory + autonomous skill creation + cross-session recall. Over time the agent accumulates a library of its own solved-problem shortcuts, compatible with the agentskills.io ecosystem.
  • 47+ built-in tools, all toggleable per deployment. Keeps the tool surface lean where it matters (safety-sensitive envs) and wide where it doesn’t.
  • 6 sandbox backends: local, Docker, SSH, Daytona, Singularity, Modal. Lets the same agent code run from a laptop up to an HPC cluster or a serverless Modal stack.
  • 15+ messaging platforms out of the box: Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI, and more. Chat surface is a connector config, not a rewrite.
  • Delegation via isolated subagents with their own conversations, terminals, and Python RPC scripts — closer to the Claude Code Agent Teams pattern than to flat function-call “sub-routines.”
  • Natural-language cron for scheduled runs (reports, backups, briefings) — the agent can be told “summarize yesterday’s inbox every weekday at 8am” without a cron expression.
  • Model-agnostic. Works with Nous Portal, OpenRouter, OpenAI, or any OpenAI-compatible endpoint. Swap providers without reconfiguring the agent logic.
  • MCP server integration — Hermes is compatible with the MCP ecosystem, so the same servers you’d plug into Claude plug into Hermes.

Architectural Pattern

Hermes sits in the “long-running personal agent” quadrant of the agent hierarchy:

  • vs Claude Code Routines — both run autonomously on their own schedule. Routines are Anthropic-hosted, Anthropic-metered, and Claude-specific. Hermes is self-hosted, model-agnostic, and carries persistent memory between runs (Routines are stateless per run).
  • vs Claude Managed Agents — both hosted-agent alternatives. Managed Agents is the turnkey Anthropic SaaS. Hermes is the open-source, BYO-infra equivalent.
  • vs OpenClaw / Claude Code Subagents — Claude Code’s subagents are ephemeral per session. Hermes subagents are spawned within a persistent parent and can be long-running themselves.

Design tradeoff: Hermes optimizes for agent longevity and memory continuity; Claude-hosted options optimize for zero-ops setup and policy guarantees.

Implementation

Tool/Service: Hermes Agent (open source, Nous Research)

Setup:

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup

Configure a model provider (Nous Portal, OpenRouter, OpenAI, or custom endpoint), select desired messaging connectors (Telegram/Discord/Slack/etc.), pick a sandbox backend (Docker for most; Modal for serverless; SSH for remote hosts), and enable the tool set you want.

Cost:

  • Runtime: free — MIT-licensed, self-hosted (your own VPS, home server, or cloud instance)
  • Inference: pay the underlying model provider directly (Nous Portal, OpenRouter, OpenAI, etc.)
  • Compare against Claude Managed Agents at $0.08/hr runtime + tokens for a hosted equivalent

Integration notes:

  • Skills format is compatible with agentskills.io, so skills built for Hermes are portable to other agentskills-compliant runtimes
  • MCP server support means existing MCP integrations (GitHub, Linear, Slack, etc.) work directly
  • Multi-provider model routing is useful for cost-tuning — route cheap tasks to smaller models, escalate complex reasoning

Try It

  1. Install on a VPS or spare machine (curl ... | bash) and run hermes setup.
  2. Point it at OpenRouter with a $5 budget cap and enable the Telegram connector so you can chat with the agent from your phone.
  3. Install one or two agentskills.io skills to seed the memory; let it auto-generate follow-ups from real work.
  4. Schedule a natural-language cron: “At 9am every weekday, summarize yesterday’s inbox and post to briefing.” Compare the output after 2 weeks vs the first day — the learning-loop claim is testable.
  5. Compare against Claude Code Routines running the same scheduled job: measure output quality, context re-use, and operational overhead.
  6. Pair Hermes with your existing MCP servers (GitHub, Sentry, PostgreSQL) so the agent can reach the same tools you use in Claude Code.

Open Questions

  • Homepage lists “five backends” for sandboxing (local, Docker, SSH, Singularity, Modal) while the docs page lists “six terminal backends” (adds Daytona). Which is current?
  • What are the concrete memory semantics — is memory a single append-only log, topic-scoped, or semantically indexed? The homepage says “never forgets how it solved a problem” but doesn’t detail retrieval.
  • How does Hermes handle model-routing decisions — explicit per-tool config, or dynamic based on task complexity?
  • What does the agentskills.io compatibility contract actually enforce? Is a Hermes skill drop-in on Claude Code Skills, or is there a translation layer?
  • Is there a hosted Nous Portal equivalent of “Managed Agents” for users who don’t want to self-host?

Graph View

Backlinks