What changed in OpenClaw 2.0?

OpenClaw 2.0 is release v2026.8.1 (GitHub release August 31, 2026). It adds a guided installer that detects existing subscriptions, API keys, and locally installed models; a rebuilt chat-first browser UI with session groups, transcript search, and fork; shared sessions with read-only, suggest, draft, and full-participation access levels; in-process GGUF inference and a managed llama-server; and a security model with per-operation permissions, Docker/Podman sandboxing, an optional 1Password broker, and plugin provenance display. Sessions and transcripts move from files to SQLite, which is a one-way migration.

Is Hermes Agent open source, and can it run local models?

Yes. Hermes Agent is MIT-licensed at github.com/NousResearch/hermes-agent. Each bot in Bot Mode is a profile under ~/.hermes/profiles/ / with its own chat, memory, skills, credentials, and a pinned model — frontier, free, or local. Providers include Nous Portal, OpenRouter, OpenAI, and your own endpoint, so a vLLM or llama.cpp server on your LAN works as a first-class model backend.

How is Grok Bot different from OpenClaw or Hermes Agent?

xAI's Grok Bot (August 11, 2026) is proprietary and hosted: each agent gets its own cloud computer, signs into your tools, and runs multi-step jobs inside xAI's infrastructure. It is bundled with Cursor Ultra, Cursor Teams Premium, and SuperGrok Heavy. OpenClaw and Hermes Agent are open source, run on your hardware, and let you pin any model — including open-weight models served on-prem — so agent state, credentials, and transcripts never leave your boundary.

Which self-hosted agent framework should a regulated organization choose?

Either can be deployed sovereignly. Choose OpenClaw 2.0 if you want the more mature permission and sandboxing model and a browser control plane; choose Hermes Agent if you want multi-bot orchestration with per-bot pinned models under a plain MIT license. In both cases lock down egress at the host, broker credentials rather than pasting keys, scope tool permissions per operation, and ship audit logs to your own SIEM. Back up state before any upgrade — OpenClaw's SQLite migration cannot be reversed without an archived transcript restore.

OpenClaw 2.0 vs Hermes Agent: Self-Hosted AI Agents in 2026

OpenClaw 2.0 and Hermes Agent v0.21.0 shipped within a day of each other at the end of August 2026, and together they define what a self-hosted AI agent framework looks like today. OpenClaw 2.0 (v2026.8.1, blog post August 30, GitHub release August 31) brings a guided installer that detects locally installed models, a rebuilt chat-first browser UI, shared sessions, in-process GGUF inference, a real permission and sandboxing model, and a one-way migration of sessions and transcripts into SQLite. Hermes Agent v0.21.0, the "Pantheon" release (August 31), is Nous Research's MIT-licensed answer: Bot Mode turns the session list into a roster of named bots, each a profile with its own memory, credentials, and pinned model, coordinated through an agent inbox and hermes peer messaging. Both stand in contrast to xAI's Grok Bot (August 11), a proprietary always-on agent whose state lives in xAI's cloud. For a regulated organization the choice between the two is secondary. The primary fact is that either one runs entirely on your own hardware against an open-weight model, and Grok Bot cannot.

What actually changed in OpenClaw 2.0?

OpenClaw's own account is that "2.0" was not planned. The team set out to simplify installation and rebuild the browser app, and carrying that cleanup through the codebase produced a major version over a roughly seven-week cycle, against a prior pace of 106 releases in 230 days. The release notes count 16,977 PRs and 987 contributors (the blog post says 933), roughly half of all PRs ever merged.

The features that matter for a private deployment:

Guided install that finds your local models. First run detects existing ChatGPT/Codex and Claude subscriptions, API keys, and locally installed models; most configuration then happens in conversation with the agent. The default model for a new install is OpenAI GPT-5.6, which a sovereign build overrides on day one.
In-process local inference. OpenClaw 2.0 runs GGUF models in-process, manages a llama-server for you, and improves the Ollama and LM Studio integrations. Hosted providers include Anthropic, DeepSeek, Cerebras, Together, Venice, OpenRouter, and Gemini.
Rebuilt chat-first Control UI. Resizable sidebar, session groups, batch actions, transcript search, and session fork. The old overview page is gone. Initial JavaScript requests drop from 140 to 45 and startup lands at 575 ms (Decrypt).
Shared sessions. Described as "multiplayer" cloud sessions with read-only, suggest, draft, and full-participation access levels. Regulated teams should confirm where shared-session state is stored before enabling it.
Channels. WhatsApp, Telegram, Discord, Slack, Signal, and iMessage, with structured agent questions delivered as cards or messaging buttons.
Storage moves to SQLite. Sessions and transcripts are no longer flat files. This is the change that bites upgraders.

What is the OpenClaw 2.0 security model?

This is where 2.0 closes the gap that made early OpenClaw installs a liability. Network installs without authentication are blocked. Remote Gateway password auth is part of onboarding, and credential prompts are masked. Automation permissions are granted per operation and can be inspected and revoked. Credentials can be brokered through an optional 1Password integration rather than pasted into config. Plugins display their provenance. The filesystem is workspace-restricted, and agent execution can run inside Docker or Podman sandboxes. These are the controls to show an auditor.

What breaks when you upgrade to OpenClaw 2.0?

The SQLite migration is a one-way door. The release notes are explicit: create a verified backup before upgrading, and before downgrading to an older file-backed release, use the current CLI to restore archived legacy transcript artifacts. Sessions created after migration will not appear if you roll back without that step. Other breaking changes: the bundled OpenProse plugin and /prose command are removed, and model references under codex/* and openai-codex/* migrate to openai/*. Both are handled by openclaw doctor --fix. InfoQ reports broken gateways, lost automations, and model-auth failures among some upgraders — reason enough to stage the upgrade.

Install and upgrade commands, per the OpenClaw docs:

# macOS / Linux / WSL2 install (Node 22.22.3+, 24.15+, or 25.9+; Node 26 recommended)
curl -fsSL https://openclaw.ai/install.sh | bash

# Upgrade an existing install, then verify
openclaw update --channel stable
openclaw --version
openclaw doctor
openclaw gateway status

Windows uses iwr -useb https://openclaw.ai/install.ps1 | iex. A sovereign deployment never pipes a remote script into a shell; mirror the installer and Node runtime inside the boundary, as our air-gapped LLM deployment guide describes.

How does Hermes Agent v0.21.0 answer?

Hermes Agent's response arrived fast. xAI launched Grok Bot on August 11. On August 14 Nous co-founder Teknium shipped Bot Mode as a public beta plugin, and on August 17 it was bundled default-on in Hermes Desktop v0.20.3. The v0.21.0 "Pantheon" release on August 31 (GitHub; tao.media dates it September 1) builds Bot Mode into the desktop app.

Where OpenClaw is one agent with many sessions and shared access, Hermes is many agents:

Bots are profiles. Each named bot lives at ~/.hermes/profiles/<name>/ with its own chat history, memory, skills, credentials, avatar, soul.md personality file, and a pinned model — "frontier, free or local." Bots can be cloned.
Agent inbox and hermes peer. Bots message each other through a persistent inbox via @mention. Group chats coordinate two to six bots across up to three serial rounds. v0.21.0 adds hermes peer bot-to-bot messaging and live subagent steering.
Cron with memory. Scheduled jobs run with persistent memory (continuity=true), so a nightly review bot remembers last night.
Unified MCP command center. v0.20.6 (August 27) added 50+ vendor MCP servers and OS-keychain secret encryption; v0.21.0 consolidates MCP management and lets the agent drive the desktop browser.
Governance gate. The agent must obtain approval before editing AGENTS.md, skills, or memory.
Providers and channels. Nous Portal, OpenRouter, OpenAI, and "your own endpoint," plus six new providers in v0.21.0. Gateways to Telegram, Discord, Slack, WhatsApp, Signal, and the CLI.

Hermes Agent is MIT-licensed at github.com/NousResearch/hermes-agent. Install:

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

Windows: iex (irm https://hermes-agent.nousresearch.com/install.ps1). The same mirror-first rule applies.

No new Hermes model shipped in this window; Nous's latest model post remains Hermes 4.3 (36B). Hermes Agent is a framework, and the model is your choice.

Why is Grok Bot the foil?

xAI's Grok Bot, launched August 11, 2026, is "a team of always-on AI agents that get their own cloud computer, sign into a customer's existing tools, and finish multi-step jobs without being supervised." It is bundled with Cursor Ultra ($200/mo), Cursor Teams Premium ($120/seat/mo), and SuperGrok Heavy. Launch coverage does not name the underlying Grok model.

The architecture is the point. Every one of those agents runs on a computer xAI operates, signed into your systems, with transcripts and credentials in xAI's cloud — the arrangement US v. Heppner made uncomfortable for law firms and OSFI Guideline E-23 makes a model-risk question for Canadian financial institutions.

There is a bridge. On May 15, 2026, xAI announced "Connect Grok to Hermes Agent": Grok subscribers can use Grok 4.3, Grok text-to-speech, and Grok Imagine inside Hermes Agent via browser OAuth, on every tier, with no API key. A team can start on Hermes with a Grok subscription and later re-pin each bot to an on-prem model without changing the agent framework. Hosted to sovereign is a config change, not a rewrite.

OpenClaw 2.0 vs Hermes Agent vs Grok Bot: comparison

	OpenClaw 2.0 (v2026.8.1)	Hermes Agent v0.21.0	xAI Grok Bot
License	Open source; nonprofit foundation since July 2026 (verify license file before procurement)	MIT	Proprietary
Storage model	Sessions and transcripts in SQLite (migrated from files; one-way)	Per-bot profile directories under `~/.hermes/profiles/`	Per-agent cloud computer in xAI infrastructure
Local-model support	In-process GGUF, managed `llama-server`, Ollama, LM Studio, llama.cpp; installer detects local models	Any model pinned per bot via "your own endpoint" (vLLM, llama.cpp), Nous Portal, OpenRouter, OpenAI	Grok only; model not stated
Multi-agent	Single agent; shared sessions with 4 access levels, session groups, event-triggered actions	Bot roster, agent inbox, `hermes peer`, group chats of 2–6 bots, live subagent steering	"Team" of always-on agents
Channels	WhatsApp, Telegram, Discord, Slack, Signal, iMessage, browser UI	Telegram, Discord, Slack, WhatsApp, Signal, CLI, desktop	Desktop (Linux), iOS; Android pending
Sandboxing and permissions	Docker/Podman sandboxes, workspace-restricted filesystem, per-operation revocable permissions, 1Password broker, plugin provenance	Approval gate on `AGENTS.md`/skills/memory edits, OS-keychain secret encryption	Managed by xAI
Where state lives	Your host; shared sessions described as cloud — confirm before enabling	Your host (`~/.hermes`)	xAI cloud
Default model on install	OpenAI GPT-5.6 (override)	None imposed; pin per bot	Grok

How do you deploy either framework sovereignly?

Both frameworks speak to an OpenAI-compatible endpoint, so the model layer is the same whichever agent you choose. August 2026 gave regulated organizations three strong open-weight options behind vLLM or SGLang:

Model	Params (total / active)	License	Context	On-prem footprint
GLM-5.3-Flash	320B / 18B	MIT	1M	~200 GB at 4-bit GGUF; ~306 GiB FP8
Qwen3.8-Flash-Next	125B / 6B (+51B N-gram table in host RAM)	Qwen Community 1.0	256K native, 1M with YaRN	172.78 GiB FP8; 111 GB at Q4
DeepSeek V4-Flash (Vision-Exp weights Aug 31)	~305B / 13B	MIT	1M	~168 GB FP8/FP4

GLM-5.3-Flash is the default recommendation: MIT, natively multimodal, and fast to decode at 18B active. Qwen3.8-Flash-Next is the pick when GPU memory is the constraint, since its N-gram table is designed to live in system RAM. Read the Qwen Community License before commercial redistribution; it is not Apache 2.0.

What a regulated organization must lock down, whichever framework it picks:

Egress. Deny-all egress on the agent host with alarmed exceptions. Neither framework needs internet access once the installer, Node runtime, and weights are mirrored inside. Both default to reaching hosted APIs; a sovereign build proves at the firewall that they do not.
Credential brokering. No API keys or service passwords in config files or transcripts. OpenClaw's 1Password broker or Hermes's OS-keychain encryption is the floor; a secrets manager your security team already audits is the ceiling.
Tool permissions. Turn on OpenClaw's per-operation permissions and Docker/Podman sandboxes for every automation. In Hermes, scope each bot's skills and credentials to its job.
Audit logs. Every tool call, model call, and permission grant goes to append-only storage you control and into your SIEM. OpenClaw's SQLite store and Hermes's profile directories are raw material, not the audit trail.
Upgrade discipline. Stage every release. Back up ~/.openclaw and ~/.hermes first, and treat OpenClaw's SQLite migration as irreversible in production.

For a law firm, the practical shape is a Hermes bot roster or an OpenClaw gateway on one on-prem node, pinned to GLM-5.3-Flash behind vLLM, with matter-level access control and no path to any vendor cloud; our law-firm page maps that to privilege and confidentiality obligations. For teams that want the agents without the operational load, our private AI hosting and maintenance practice runs the model layer, the agent framework, and the upgrade cadence on hardware you own. Reference architectures and a sovereign-deployment checklist are in our resources library.

OpenClaw 2.0 and Hermes Agent v0.21.0 make the always-on agent something you can run on your own rack. The differentiator is no longer whether an agent can work unsupervised. It is whether it does so inside your boundary.