Hermes Agent Review 2026: Features, Performance, and Honest Assessment

Hermes Agent is an open-source, MIT-licensed AI agent runtime built by Nous Research. It runs on your own infrastructure, connects to LLM providers of your choice, and executes tasks using 70+ built-in tools. The framework launched on February 25, 2026, and reached version 0.13.0 (the "Tenacity Release") on May 7, 2026.

The pitch is straightforward: Hermes is the agent that learns from you. Unlike most agent frameworks where every session starts fresh, Hermes maintains persistent memory across conversations, auto-generates reusable skills from completed tasks, and builds a user profile that reduces repeated context-steering over time. After 20+ accumulated skills in a domain, similar tasks complete roughly 40% faster than they would on a clean instance.

That learning loop is what separates Hermes from OpenClaw, Claude Code, and other agent runtimes. OpenClaw has a larger skill ecosystem (44,000+ community skills on ClawHub) and broader messaging coverage. Claude Code is purpose-built for software engineering. Hermes occupies a different niche: a general-purpose agent that compounds capability the longer you use it.

The framework supports 20+ messaging platforms (Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Microsoft Teams, Email, SMS, and more), six deployment backends (local, Docker, SSH, Daytona, Modal, and Singularity), and native MCP client/server support. It is model-agnostic, working with Nous Portal, OpenRouter, OpenAI, Anthropic, or any custom endpoint.

Version 0.13.0 shipped with 864 commits, 588 merged pull requests, 282 closed issues, and contributions from 295 developers. That velocity is unusual for a project under four months old.

Most Hermes reviews describe the memory system in abstract terms. Here is how it actually works.

Hermes uses a three-layer memory architecture. The first layer is identity snapshots stored in two Markdown files: MEMORY.md (agent notes, roughly 800 tokens) and USER.md (user profile, roughly 500 tokens). These get injected into the system prompt at session start, giving the agent immediate context about who you are and what you care about. The second layer is a SQLite database with FTS5 full-text search indexing every conversation. Retrieval latency sits around 10ms even across 10,000+ stored documents. The third layer is Honcho dialectic user modeling, which builds a progressively deeper understanding of individual users based on interaction patterns.

The skill generation side works like this: when Hermes completes a task involving five or more tool invocations, it pauses, reflects on what worked, and writes a reusable skill file following the agentskills.io open standard. The next time a similar task appears, the agent loads the existing skill instead of reasoning from scratch. Skills are portable Markdown files you can inspect, edit, and share across Hermes instances.

What works well: After several weeks of daily use, the agent genuinely stops asking for context you have already provided. If you start a new image task, it already knows your brand guidelines, preferred tool stack, and output preferences. The 40% speed improvement on domain-specific tasks is real, but it requires accumulating enough skills in that specific domain (typically 20+).

Where it breaks down: The identity snapshot files are tiny. MEMORY.md caps out at roughly 2,200 characters and USER.md at roughly 1,375 characters, which adds up to about 20 short entries total. Hit that ceiling and the agent spends turns consolidating and replacing entries instead of doing useful work. The SQLite archive is thorough but not easily human-readable. Unlike OpenClaw's transparent file-per-memory approach, you cannot easily export everything Hermes knows about you into a format you can audit. For personal use this is a minor annoyance. For anything regulated (healthcare, finance, legal), this opacity creates real compliance friction.

The critical gotcha: Self-learning is disabled by default. You need to explicitly enable persistent_memory and skill_generation in configuration. A surprising number of early reviewers dismissed Hermes as unremarkable because they never flipped these switches.

Hermes ships with 70+ built-in tools covering web search, content extraction, browser automation, vision analysis, image generation, text-to-speech, and file operations. Version 0.13.0 added video_analyze for native video understanding on Gemini-compatible models and post-write delta linting for Python, JSON, YAML, and TOML files.

The skill ecosystem includes 118 bundled skills curated by Nous Research, plus a growing registry on agentskills.io. Six new optional skills arrived in 0.13.0: Shopify integration, here.now, shop-app, Anthropic financial-services, kanban-video-orchestrator, and searxng-search. Compared to OpenClaw's 44,000+ community skills on ClawHub, Hermes trades breadth for quality. Nous Research scans every bundled skill for security issues, which matters given that a Koi Security audit found an 11.9% malware rate in ClawHub submissions.

On messaging, Hermes supports Telegram, Discord, Slack, WhatsApp, Signal, DingTalk, SMS (via Twilio), Mattermost, Matrix, Email (IMAP/SMTP), Home Assistant, Feishu/Lark, WeCom, Weixin, BlueBubbles (iMessage proxy), QQBot, Yuanbao, Microsoft Teams, and Google Chat (added in 0.13.0). A unified session layer means you can start a conversation in CLI, continue it in Telegram, and route the result to your team in Slack. All platforms share one session and one memory store.

Voice features work across CLI (push-to-talk), Telegram (voice notes), and Discord (voice channel support), with local Whisper transcription via faster-whisper when the voice extra is installed.

Real limitation: The token overhead is significant. Community analysis found that roughly 73% of each API call is fixed overhead, with tool definitions consuming about 46% (around 8,759 tokens) and the system prompt taking another 27% (around 5,176 tokens). Only about 27% of each call is your actual conversation. Through messaging gateways, overhead climbs to 15,000-20,000 input tokens per request. You can mitigate this with selective tool loading and model routing, but it is a real cost factor.

Persist Hermes Agent output across sessions and teams

Free 50GB workspace with MCP endpoint your Hermes instance can write to directly. No credit card, no trial expiration, no setup friction.

Sign Up Free

Hermes runs on six terminal backends: local (Linux, macOS, WSL2, Android via Termux), Docker, SSH, Daytona, Modal, and Singularity. Daytona and Modal offer serverless persistence with near-zero idle costs. A basic always-on setup runs fine on a $5/month VPS.

The security picture is mixed but trending well. Hermes had zero agent-specific CVEs through April 2026, which compares favorably to OpenClaw's nine CVEs disclosed in a four-day window in March 2026 (including one scoring CVSS 9.9). However, less exposure time does not automatically mean better security engineering. Version 0.13.0 addressed eight priority-zero security issues: secret redaction is now enabled by default, Discord role-allowlists are guild-scoped (closing a CVSS 8.1 cross-guild DM bypass), WhatsApp rejects messages from unknown contacts by default, and TOCTOU windows in auth.json and MCP OAuth flows are closed. Browser operations enforce cloud-metadata SSRF mitigations, and cron jobs scan for prompt injection in assembled skill content.

The container security model includes read-only root filesystems, dropped capabilities, namespace isolation, and filesystem checkpoints with rollback. A pre-execution scanner reviews terminal commands before they run. These are reasonable defaults for a self-hosted agent.

Production verdict: Hermes is production-viable for solo developers and small teams running specific, repeatable workflows over extended periods (six months or more). It is not production-ready for enterprise customer-facing systems. API stability between v0.x releases is not guaranteed, the codebase is under four months old, and documentation for community-built extensions (PLUR engrams, hermes-workspace, mission-control) is often incomplete. Pin your version and test upgrades before deploying them.

Hermes itself is free under MIT. You pay for LLM API calls, hosting, and optionally for external services your skills use. Typical cost breakdown for moderate personal use: $30-65/month in API calls on a $5/month VPS. Heavy autonomous workflows can run $800-1,500/month in API costs. Budget models bring individual complex tasks down to roughly $0.30 each.

Strengths worth highlighting:

• The learning loop is genuinely novel. No other open-source agent framework attempts cross-session self-improvement at this level.
• Security posture is strong for the category. Curated skills, default redaction, and container hardening are real differentiators.
• Model-agnostic design means you are not locked into a specific LLM provider. Use Nous models, Claude, GPT-4, Gemini, LLaMA, or local models interchangeably.
• MCP support works both ways. Hermes connects to external MCP servers as a client and exposes its own conversations as an MCP server for other agents.
• The /goal command in 0.13.0 keeps the agent focused on a target across conversational turns. This is surprisingly useful for complex, multi-step tasks that other agents lose track of.

Weaknesses to consider:

• Setup complexity is real. You need to choose a provider, manage API keys, configure tools, and enable self-learning manually. This is not a download-and-go experience.
• Memory limits are frustratingly small. Hitting the 2,200-character MEMORY.md ceiling after a few weeks of use forces the agent into consolidation loops instead of productive work.
• Not a coding copilot. Hermes underperforms compared to Cursor and Claude Code for software engineering tasks. It has no inline autocomplete, no IDE integration, and no specialized code generation pipeline.
• The token overhead tax means you pay for roughly 12,000-16,000 tokens of fixed context before the agent processes a single word of your message.
• Documentation gaps exist. Core docs are solid, but community-built features often require reading GitHub issues for setup instructions.

Best fit: Solo developers and researchers who interact with their agent daily over months, building up domain-specific skills in areas like research, content creation, monitoring, and automation. If you need broad team deployment, start with OpenClaw. If you need a coding assistant, use Cursor or Claude Code. If you want an agent that gets measurably better at your specific workflows over time, Hermes is the strongest option in 2026.

One gap in the Hermes workflow is output permanence. The agent generates files, research summaries, reports, and processed data, but all of it lives on whatever server Hermes runs on. When the agent finishes a task, someone still needs to move the output somewhere accessible, share it with collaborators, or hand it off to a client.

Local storage works for solo use. For anything involving collaboration or handoff, you need a layer between the agent and the people who consume its output. Options include S3 buckets (cheap, no collaboration features), Google Drive (familiar but limited API tooling for agents), or a purpose-built workspace like Fast.io that is designed for agent-to-human handoff.

Fast.io fits this gap specifically. The free agent plan includes 50GB of storage, 5,000 credits per month, five workspaces, and no credit card requirement. The MCP server exposes workspace, storage, AI, and workflow operations that Hermes can call directly through its native MCP client. Intelligence Mode auto-indexes uploaded files for semantic search and RAG, so team members can ask questions about agent output without digging through folders.

The practical workflow: Hermes generates output on its server, uploads to a Fast.io workspace via MCP, and the workspace owner (a human) reviews, shares, or transfers it. Ownership transfer lets an agent build out a workspace and hand full control to a human while retaining admin access for future updates. This matters because Hermes excels at building up domain-specific knowledge over months, but the value of that knowledge only compounds when other people can access and act on it.

You can test this integration on the free plan without any commitment. The MCP endpoint works with Hermes' native MCP client configuration.