Hermes Agent
A self-hosted, multi-channel agent that builds skills and retains context across sessions.
Per-dimension scores and reasoning
Evidence shows a security policy document (SECURITY.md) detailing trust model, isolation boundaries, and credential scoping, but no actual execution verification. Dependency security: pyproject.toml pins exact versions with CVE comments, but no vulnerability scan results. External effects: documentation mentions command approval and sandboxing, but no actual tests. Rollback: documentation mentions /undo command, but no implementation details. Source attribution: documentation clearly attributes authorship to Nous Research, but not verified.
Self-consistency: documentation and code comments are consistent, but no runtime verification. Dependency availability: dependency list is complete, but no availability tests. Failure messages: documentation mentions diagnostic commands, but no specific error handling examples.
Audience and scenarios: documentation describes multiple use cases, but no concrete examples. Capability boundaries: documentation lists tools and platforms, but no explicit limitations. Trigger precision: documentation mentions commands and skills, but no precise trigger conditions. Environment fit: documentation mentions multiple deployment methods, but no configuration details.
Information architecture: documentation structure is clear, but no complete directory. Install notes: installation commands provided, but no detailed steps. Naming stability: project name and commands are consistent, but no version history. Examples and FAQ: basic examples provided, but no FAQ. Known limitations: documentation mentions some limitations, but not comprehensive. License: MIT license is clear. Versioning and changelog: pyproject.toml has version number, but no changelog. Maintenance responsibility: documentation mentions contribution guide, but no explicit maintainers.
Output usability: documentation describes output formats, but no actual output examples. Marginal value: documentation emphasizes unique features, but no comparison. Cost-benefit: documentation mentions low-cost deployment, but no specific cost analysis.
Claim traceability: claims in documentation lack specific evidence. Cross-source corroboration: no external verification. Fact-inference separation: documentation does not clearly distinguish facts from inferences.
- Static review cannot verify actual security and reliability; runtime testing is needed.
- Dependency security is pinned but no vulnerability scan results are provided; check CI OSV scan.
- Feature claims in documentation lack specific evidence; treat with caution.
What does this agent do, and when should you use it?
Hermes Agent is Nous Research's self-improving agent with an interactive CLI/TUI and a unified messaging gateway. It supports conversations through CLI, Telegram, Discord, Slack, WhatsApp, Signal, and Email-oriented gateway documentation, with the gateway launched through `hermes gateway`. Its learning loop includes skill creation, skill improvement during use, persistent memory, FTS5 session search with LLM summarization, and user modeling. Work can run through local, Docker, SSH, Singularity, Modal, Daytona, or Vercel Sandbox terminal backends, alongside cron scheduling and isolated subagents. The model layer can be switched among Nous Portal, OpenRouter, OpenAI, self-hosted endpoints, and other supported providers.
Run hermes for a terminal interface with multiline editing, command autocomplete, conversation history, interruption, and streamed tool output; or use hermes gateway setup followed by hermes gateway start to connect messaging platforms. hermes model selects a provider and model, hermes tools configures enabled tools, and hermes config set and hermes config get manage individual configuration values. The agent creates skills, improves them during use, and recalls earlier work through persistent memory, FTS5 session search, and LLM summarization. It can work through seven named terminal backends, spawn isolated subagents, use Python scripts that call tools via RPC, and schedule unattended work with its built-in cron scheduler.
- A solo developer who wants to monitor work on a cloud VM from Telegram or Slack instead of staying at one terminal.
- An operations engineer who needs to choose between local, Docker, or SSH execution while using an agent that can run shell commands.
- A small team scheduling daily reports, nightly backups, or weekly audits and delivering results through a messaging channel.
- A knowledge worker who wants recurring work to produce reusable skills and retrieve relevant prior-session context.
- An OpenClaw user who needs to import a persona, memories, skills, messaging settings, or supported API keys.
What are this agent's strengths and limitations?
- Combines a terminal UI with gateways for Telegram, Discord, Slack, WhatsApp, and Signal, with documented cross-platform conversation continuity.
- Its documented learning loop includes skill creation, in-use skill improvement, memory nudges, FTS5 search, and LLM summarization.
- Seven terminal backends span local execution, containers, remote SSH, and serverless-style environments.
- Built-in cron scheduling, isolated subagents, and RPC-based tool scripts support unattended multi-step workflows.
- Model, messaging, and tool integrations still require configuration and credentials; the supplied material does not enumerate every provider's key setup procedure.
- The documented installation brings runtime dependencies including Python 3.11, Node.js, ripgrep, and ffmpeg; Windows also needs Git or the installer-managed MinGit.
- Termux uses a curated
.[termux]extra rather than the full.[all]set because some voice dependencies are currently Android-incompatible. - OpenClaw migration may import API keys, messaging settings, and workspace instructions, so
--overwriteshould be used carefully.
How do you install or deploy this agent?
On Linux, macOS, WSL2, or Termux, run curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash. On native Windows PowerShell, run iex (irm https://hermes-agent.nousresearch.com/install.ps1). After installation, run source ~/.bashrc (or source ~/.zshrc), then run hermes to start a conversation. The installer is documented as handling uv, Python 3.11, Node.js, ripgrep, and ffmpeg; on Windows it uses an existing Git installation when found or installs an isolated MinGit copy.
How do you use this agent?
Start with hermes, then use hermes model to select a model provider and model; the supplied material does not specify the credential setup commands for every provider. For Nous Portal, run hermes setup --portal; this signs in through OAuth, selects Nous as provider, and enables the Tool Gateway. To use a messaging platform, run hermes gateway setup and hermes gateway start, then message the bot. Within a conversation, use commands such as /new, /model [provider:model], /skills, /compress, and /usage; use hermes doctor for diagnostics.
How does this agent compare with similar options?
The repository names OpenClaw as a migration source. hermes claw migrate can import SOUL.md, memories, skills, command approval patterns, messaging settings, selected API keys, TTS assets, and workspace instructions; --dry-run previews the migration first.
FAQ
Can I use a provider other than Nous Portal?
hermes model.Is Hermes limited to a laptop?
Can it run scheduled tasks and send their results somewhere?
What should I watch for when migrating from OpenClaw?
hermes claw migrate --dry-run first and use --overwrite only after reviewing conflicts.