Hermes Agent

Every entry point (CLI/TUI, messaging gateway, ACP adapter, API server, batch runner) drives the same AIAgent loop, which builds the prompt, resolves the provider and dispatches tools from a registry. Sessions are stored in SQLite with FTS5 search, and tools run on one of several terminal backends (local, Docker, SSH, Modal, Daytona and others).

In-process multi-agent work is supervisor-style. The model calls delegate_task with a goal and context; each child starts with a fresh conversation (plus the workspace's AGENTS.md/CLAUDE.md context files), inherits the parent's toolsets, and returns only a summary. Top-level delegation runs in the background and posts results back as a new message; children are leaf workers by default, and orchestrator children can delegate further only when delegation.max_spawn_depth is raised.

For longer-lived collaboration, the Kanban board stores tasks and handoffs in ~/.hermes/kanban.db, and a dispatcher starts each worker as a full OS process under its own profile. Across machines or frameworks, the A2A plugin connects independent agents. Memory is two capped Markdown files (MEMORY.md and USER.md) per profile, loaded as a frozen snapshot at session start, with optional external memory providers.

Dangerous shell commands go through approvals.mode: smart (default; an auxiliary LLM auto-approves low-risk commands, denies clearly dangerous ones and escalates the rest), manual (always prompt) or off. In the CLI the prompt offers once, session, always or deny; on messaging platforms the user replies yes or no in chat. /yolo or --yolo bypasses prompts for a session. /stop ends a run and its background subagents, and the CLI, TUI and Desktop show live subagents that can be steered or stopped individually. Subagents cannot use the clarify tool, so they never ask the user directly. Unknown messaging users must be approved with hermes pairing approve.

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

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
source ~/.bashrc          # or ~/.zshrc

hermes setup              # provider, tools, optional messaging gateway
hermes model              # pick provider and model

# allow one level of orchestrator subagents and cap parallel children
hermes config set delegation.max_spawn_depth 2
hermes config set delegation.max_concurrent_children 4

# let CLI sessions call remote A2A agents (toolset is off by default)
hermes tools enable a2a --platform cli

hermes                    # then ask it to research two topics in parallel

• The POSIX installer needs Git, curl, tar and SHA-256 utilities; first-party installs run on Python 3.14 and the installer pins its own uv.
• On Android/Termux use the signed APT repository, not the install script. On native Windows use the PowerShell one-liner; some antivirus tools flag the bundled uv.exe as a false positive.
• Subagents start with no knowledge of the parent conversation; everything they need must go in goal and context.
• Raising delegation.max_spawn_depth multiplies concurrent agents and cost (depth 3 with 3 children each can reach 27 leaf agents).
• Do not run two agent processes against the same Hermes home directory; memory writes from both end up in each other's prompts.

Yes. It connects to MCP servers as a client and can run as an MCP server with hermes mcp serve. Its A2A plugin both serves tasks to other agents and calls remote A2A agents. AG-UI support is not documented.

The code is MIT-licensed and works with your own provider keys. Nous Research also sells Nous Portal subscriptions, which bundle model access and a managed Tool Gateway for Hermes; they are optional.

Three ways: delegate_task subagents inside one process, a Kanban board that coordinates several Hermes profiles as separate processes, and A2A for agents on other machines or frameworks.

The README documents hermes claw migrate, and hermes setup offers to import settings, memories, skills and API keys when it finds ~/.openclaw.

Python; first-party installs run on Python 3.14, which the installer provisions.