Hermes Agent alternatives
Hermes Agent is free and open source, with an optional paid Nous Portal subscription for access to 300-plus hosted models. It's built by a venture-backed company, Nous Research, which is the single biggest thing that separates it from OpenClaw, its closest and most-searched-for alternative. Both are self-hosted, and the incentives behind the company running each one differ. Mover OS is also an option if you want help following your own plan more than errands done for you: $49 once, it runs inside the coding agent you already use, such as Claude Code or Codex, keeps plain files on your own machine, plans your day, flags drift from your plan, and stops when the agent closes.
Last verified 29 September 2026.
Which alternative fits you?
Try OpenClaw
Worth a look if a venture-backed company's paid tiers sitting behind your agent bothers you. OpenClaw is stewarded by an independent nonprofit foundation with, by its own account, no paid tier, hosted service or token to ever sell.
Try Poke
A good fit if self-hosting anything, even something free, is more setup than you want. Poke works today over the texts you already send, no cloning a repo or picking a model provider required, in exchange for giving up Hermes's autonomous skill-writing and 300-plus model choice.
Try Mover OS
It suits you if what drew you to Hermes was its memory and correction-tracking (MEMORY.md, USER.md) rather than its messaging-gateway reach or autonomous skill-writing. Mover does something adjacent inside Claude Code, Codex or Gemini CLI: it tracks your plan in plain files and calls out drift. Mover is the wrong choice if you want a general-purpose agent reachable from Telegram, Discord, Slack and other chat apps: Mover doesn't connect to messaging platforms, and it doesn't run unattended cron jobs the way Hermes does.
How do the alternatives compare?
| Fact | Hermes Agent | OpenClaw | Poke |
|---|---|---|---|
| Maker | Nous Research (venture-funded; TechCrunch reported a $75M round at a $1.5B valuation, July 2026) source | The OpenClaw Foundation, an independent 501(c)(3), donation-funded; original creator Peter Steinberger source | The Interaction Company of California (reported as now part of Cognition) source |
| Launch date | Repository created 22 July 2025; public launch "since February 2026" per its own star-growth framing (64K+ stars since Feb 2026 launch, 247,436 stars by 20 September 2026) source | Repository created 24 November 2025; continuous releases since (e.g. 2026.9.5 on 19 September 2026) source | Officially launched September 2025 (with a $15M seed round); approved by Apple for native Apple Messages for Business access on 4 June 2026 source |
| Price and plans | Free and open source (MIT license) to self-host with your own model API keys. A Nous Portal subscription (for 300+ bundled models plus a Tool Gateway) is reported at Free/$20/$100/$200 per month tiers, triangulated from third-party SEO pages and TechCrunch's reported $20-$200/mo range; the official pricing page itself returned HTTP 429 to automated checks in this and the prior pass, so exact tier names and credit amounts are UNVERIFIED against the primary source. Separate Business and Enterprise tiers exist (contact sales). source | Free, open source, self-hosted. Official FAQ: "no paid tier, no hosted service, no token." A separate domain, getopenclaw.ai, advertises an unaffiliated-looking "$49/mo Cloud" offer that contradicts the official Foundation FAQ; not used here as an official price, flagged as a likely unofficial/SEO site rather than the Foundation's own product. source | Free ($0/mo, no card required); Pro $19/mo ($16/mo billed yearly); Ultra $199/mo ($160/mo billed yearly, includes pay-as-you-go usage beyond credits and the "Poke Human" add-on). Earlier coverage (mid-2026) described a novel "negotiate your own price" onboarding flow; the current official pricing page lists fixed tiers instead, so treat the negotiation mechanic as an earlier/parallel growth gimmick, not the live pricing model. source |
| Availability | Global; self-hosted, open source, no account required for the core agent (an anonymous/free-tier auth path exists for those who do want hosted models). source | Global; self-hosted on hardware you control, any OS. source | All regions worldwide, per Poke's docs. source |
| Platforms | macOS, Linux, Windows; a desktop app; 22+ first-party messaging-platform plugins (Telegram, Discord, Slack, WhatsApp, Matrix, LINE, Feishu, Teams and more) plus generic webhook/API-server ingress. source | macOS, Linux, Windows (native and WSL2), Raspberry Pi; reachable via Discord, iMessage, Signal, Slack, Telegram, WhatsApp, Matrix, IRC, Nostr and 20+ more channels, plus mobile voice and hosted widgets. source | iMessage (first AI agent Apple approved for Messages for Business), WhatsApp, Telegram, RCS/SMS. No dedicated app required. source |
| Underlying model | Model-agnostic via the Nous Portal: 300+ models including Anthropic Claude, OpenAI GPT, Google Gemini, DeepSeek, Qwen, Kimi, GLM, MiniMax, Grok, and others; or local models via Ollama/llama.cpp, subject to a minimum-context-length floor that can reject small local models by default. source | Model-agnostic: Anthropic, OpenAI, MiniMax, OpenRouter and others via plugin, with per-agent routing and failover; a local-only option runs entirely on-device models (llama.cpp, LM Studio, Ollama, MLX, vLLM, SGLang). source | A blend of models from Anthropic, OpenAI, Voyage and Mistral AI plus in-house fine-tuned models; which model handles which task is not disclosed. source |
| Where memory lives | Two local files, MEMORY.md (2,200-character budget) and USER.md (1,375-character budget), on the user's own machine under ~/.hermes/. Injected as a frozen snapshot at session start; the tool has no read action, and by its own design has no source-session lineage or deletion tombstone on entries (per a competitor's technical comparison). source | Local, on the user's own hardware, alongside session state and credentials. "State, memory, and credentials live on your hardware." A competitor's own technical comparison (OpenClaw's docs) notes memory entries carry no source-session lineage or deletion tombstones by default. source | A per-account Memory inside Poke's hosted service, kept together with your Recipes and Integrations when two accounts are merged. source |
| Always on | Not by default; a full cron/scheduler subsystem exists for unattended, scheduled jobs when self-hosted continuously, including a script-only wake-gate mode with no LLM call. source | Depends on the user's own setup: yes if the host machine (e.g. a Raspberry Pi gateway) is left running continuously, no built-in cloud always-on layer of its own. source | Yes. Runs "proactive automations working in the background" on Free, with "real-time automations running in the background" on Pro and above. source |
| Integrations | MCP; the Nous Portal Tool Gateway (Firecrawl web search/extract, FAL image generation, OpenAI TTS, Browser Use cloud browser automation, Modal cloud sandbox); autonomous skill authoring (skill_manage) and a background skill-consolidation "Curator". source | MCP client and server, the Linux Foundation A2A 1.0 agent-to-agent protocol, Agent Client Protocol (ACP) for hosting external harnesses (including Claude Code and Codex as plugins), the AgentSkills spec, an OpenAI-compatible /v1/chat/completions and /v1/responses API, OpenTelemetry and Prometheus export. source | Email and calendar (Free tier onward); Notion, Linear, Oura and roughly 20 more cited by a third-party review; Recipes and an MCP layer for power users. source |
| Trains on your data by default | No. Hermes collects no telemetry, usage data or analytics; conversations, memory and skills stay in ~/.hermes/ and model calls go only to the provider you configure. source | Not applicable in the way a hosted product has a default: it is self-hosted software with no central account. "By default OpenClaw itself phones home for nothing but a daily version check; anonymous feature statistics are opt-in." source | A third-party review states "no training on your data" as part of Poke's privacy stance; Poke's own published policy text stating a default was not independently located in this pass. source |
Each of the 30 cells in this table links to its source. A row is left out when no source answers it for every product shown.
Mover OS is a different kind of tool: files your own AI agent reads, $49 once, with no cloud computer or messaging reach. Hermes vs Mover OS sets the two side by side.
What do users say about Hermes?
“I've been looking for a new cooker. I have a Hermes agent running on VM that does this. I told it to keep an eye on prices. It created a cron job that tri[ggers]...”
cassianoleal, Hacker News, Hermes Agent
“IMO this comes down to your harness. Any frontier model from a huge shop has an inherent benefit in the system you're using it in... I recommend trying Hermes Agent and setting it up fully, that's the closest you'll get to a more complete experience.”
wronglebowski, Hacker News, Hermes Agent
“I also run Hermes Agent and have set up there one agent to be responsible for the KB stuff and that one I connected to a Matrix channel and can use it from the phone and any device where I have a matrix client like Element set up.”
jeena, Hacker News, Hermes Agent
“The local model forgot things without saying so. Every model can hold only a limited amount of text at once, its context window, and in long conversations the local model was silently dropping the beginning.”
Marc Bara, Medium, Hermes Agent
Common questions
Is Hermes Agent free?
The core agent is free and open source under the MIT license and runs on your own API keys at no extra cost beyond your model provider's bill. An optional Nous Portal subscription bundling 300-plus models and extra tools is reported at $20 to $200 a month, though the official pricing page returned an HTTP 429 error on every attempt to read it directly, so exact tiers remain unconfirmed against the primary source.
What does Hermes's autonomous skill-writing actually do?
Roughly every 10 tool-call iterations or every 10 user turns (both configurable), Hermes forks a background process that reviews the conversation and, if it finds a reusable pattern, writes or patches a skill file for future sessions automatically, without blocking your live reply. A separate, much less frequent weekly pass consolidates narrow one-session skills into broader ones so the library doesn't sprawl unbounded.
Can Hermes run entirely offline with a local model?
It supports local models through Ollama, llama.cpp and others, but it enforces a minimum context-length floor and will reject a small local model outright unless you manually override the setting, a real friction point at least one independent user reported hitting.
How is OpenClaw different if it's also self-hosted and free?
The clearest structural difference is governance and incentive: OpenClaw's own comparison page puts it directly, "Venture investors expect a return, and the reported plan for that return is paid Hermes tiers. OpenClaw's only obligation is to stay useful enough that people keep donating." Feature-wise, Hermes adds autonomous skill-writing and a full cron/scheduler subsystem that OpenClaw's core doesn't ship with in the same form.
Weighing Hermes against a tool that holds you to your own plan? Mover OS: $49 once, plain files on your own machine, inside the AI you already use.