Hermes Agent Draft
Nous Research · reviewed 22 November 2026 · its card
Hermes Agent lets one person keep a personal agent that remembers them and works unattended, reached from a terminal, a desktop app or a chat app: "the self-improving AI agent", with "a built-in learning loop" that "creates skills from experience" and "searches its own past conversations", that can run "on a $5 VPS, a GPU cluster, or serverless infrastructure" (S2). Nous Research makes it, in the NousResearch GitHub organisation (S1, S2). It runs as a terminal interface and CLI, a desktop app for macOS, Windows and Linux, a browser Webapp, a messaging gateway onto Telegram, Discord, Slack, WhatsApp, Signal, email and other platforms, and on Android through Termux (S2, S4, S14, S15). It "is free and open source under the MIT license", with models and hosted services, such as the Nous Portal subscription, priced separately (S1, S2). The latest release seen is v0.23.3, of 19 November 2026 (S3). It is an agent harness and a client; Orca runs it as a bot agent (S23).
- An instance of
- Agent harness, Client
Capabilities
| Capability | Value | Detail | Sources |
|---|---|---|---|
| Model | Yes | "Use any model you want — Nous Portal, OpenRouter, OpenAI, your own endpoint", self-hosted Ollama, vLLM and LM Studio, and local models through a managed llama.cpp; "Switch with hermes model" |
S2 S20 S21 |
| Session | Yes | "Hermes Agent automatically saves every conversation as a session"; hermes --continue and --resume by id or title |
S5 |
| Subagent | Yes | delegate_task "spawns child AIAgent instances with isolated context"; "only its final summary enters the parent's context" |
S6 |
| Agent team | Yes | Bot Mode: "Bots run routines, share group chats, and message each other", a room's lead bot can pick who answers; Kanban lets "multiple named agents collaborate" through a shared board | S7 S4 S22 |
| Skill | Yes | SKILL.md skills in ~/.hermes/skills/ and external folders, with progressive disclosure; the maker names the standard: "compatible with the agentskills.io open standard" |
S8 S2 |
| Instruction file | Yes | .hermes.md, AGENTS.override.md, AGENTS.md, CLAUDE.md and .cursorrules, first match wins; AGENTS.md merged from the git root down and found progressively in subfolders; a global SOUL.md |
S9 |
| Plugin | Yes | "a plugin system for adding custom tools, hooks, and integrations without modifying core code"; hermes plugins install from a catalogue or Git; the catalogue passed 1,050 entries in v0.23.0 |
S10 S4 |
| MCP | Yes | "Local stdio servers and remote HTTP MCP servers in the same config", with OAuth | S11 |
| Execution environment | Yes | none by default, the local terminal backend; container with Docker; vendor-sandbox with Modal, Daytona or Vercel Sandbox; SSH to another machine |
S13 S12 |
| Client | Yes | Terminal (CLI and TUI), desktop, web (the Webapp, off by default, and a dashboard), messaging apps, Android through Termux, and editors over ACP | S2 S4 S14 S15 S17 |
| Agent host | Partial | Opt-in only: OpenAI turns can be handed to the Codex CLI app-server, "Hermes becomes the shell around it (sessions DB, slash commands, gateway, memory and skill review)" | S19 |
| Channel | Yes | The messaging gateway brings chats from many platforms into its sessions; webhooks "trigger Hermes agent runs", and mirror_to_session appends a delivery to a chat's session; inbound A2A tasks are "injected into a live gateway session" |
S15 S16 S18 |
| Cross-session messaging | Yes | "every Bot Chat carries the message_agent tool", delivering "into the teammate's canonical Bot Chat"; on a headless install it must be enabled by hand |
S7 |
| ACP | Yes | As agent: "Hermes Agent can run as an ACP server, letting ACP-compatible hosts talk to Hermes over stdio", after installing the ACP extra | S17 |
| AHP | Unknown | Not found | None |
| A2A | Yes | The A2A plugin in the repository "works in both directions": an Agent Card and JSON-RPC server inbound, a2a_call and other tools outbound, the toolset off by default |
S18 |
| Bot | Yes | Bot Mode, on by default in the desktop app: "Each Bot has its own role, model, memory, skills, and avatar"; routines are cron jobs: "Runs land in the Bot's own chat history" | S7 |
State of play
How things stand, newest first, as the card records it.
- Orca v1.4.268 lets Orca bots "run on Cursor, GitHub Copilot and Hermes"; the year is from the Orca card's dating of that release. S23
- Version 0.23.3, a patch release rolling up about 370 pull requests since v0.23.2, with "full curated release notes" left for v0.24.0. S3
- Version 0.23.0, "The Hearth Release": the desktop workspace in a browser tab, "one gateway owns every conversation" with every surface attaching as a client, Group Chats that "run themselves", and pluggable delegation. S4
- Version 0.22.0, "The Threshold Release", with a guided first-run setup chat on a free trial model, connectors and one-click local models. S24