As of 2026-05-28

OpenClaw is the project that turned "fire-and-forget personal agent" from a research demo into something you could realistically run from a Discord channel. It has been the reference open-source autonomous agent of late-2025-onward, the project Anthropic explicitly responded to with Claude Code Channels, and the de-facto benchmark every newer self-hosted agent gets compared against.

This article is the working engineer's read of how OpenClaw works, where it came from, and what the design choices imply for using it on real work.

Origin and history

OpenClaw was released in November 2025 by Peter Steinberger, an Austrian developer known previously for iOS tooling work (PSPDFKit). The first release was named "Clawd" — a play on Claude, the model it was built around.

The rename to OpenClaw came after Anthropic sent a cease-and-desist citing trademark concerns with the "Clawd" name. The project rebranded, kept the "claw" phonetic identity, and continued shipping. The episode became part of the project's public identity: comparison pieces in 2026 still routinely mention the Clawd → OpenClaw rename when introducing the project.

By early 2026, adoption had grown fast enough that Anthropic shipped Claude Code Channels — a vendor-built autonomous agent with messaging integrations — explicitly aimed at the workflow OpenClaw had popularized. The Anthropic move both validated the category and put commercial pressure on it.

The core UX: messaging-first, asynchronous

The defining shape of OpenClaw is that you do not sit in front of an IDE waiting for the agent to finish. You message it.

The typical interaction model:

  1. You open a Discord channel, Telegram chat, Slack thread, or terminal session connected to your OpenClaw instance.
  2. You send a task: "Update the README to reflect the v2 API," or "Run the test suite and fix the failures," or "Triage the last 30 GitHub issues and label them."
  3. OpenClaw acknowledges, then works in the background — clones the repo if needed, runs commands, edits files, runs tests, iterates on failures.
  4. When done, it messages you back with a summary, a diff, and a link to any PR or branch it created.

This is the "fire-and-forget" model. It does not require you to babysit. You can ask it to do something at the start of a workout, get a notification when it finishes, and review the work.

The implication for what kinds of work fit: tasks that are well-specified enough that you can describe them in a single message, and that have a clear "done" signal (tests pass, PR opens, output produced). Less suited to exploratory tasks where you would naturally want to steer mid-process.

The skill system

OpenClaw exposes its capabilities through a skill abstraction. A skill is a unit of behavior the agent knows how to perform — "refactor this module," "summarize this issue," "deploy this service" — written as a structured file with prompts, tools, and expected outputs.

Skills are human-authored. Someone writes a skill, ships it to the community library, and others pull it in. The library is large by mid-2026; community comparison posts cite tens of thousands of skills across categories like coding, ops, content, research, and customer support.

The trade-off versus Hermes Agent's autonomously-generated skills:

  • Human-authored skills are more reliable for the use case they were built for. Someone debugged them, tuned the prompts, and shipped them.
  • Human-authored skills do not adapt to your specific environment without you forking and editing them.
  • The library is wider than any one team would build internally, so for general-purpose tasks the right skill probably already exists.
  • For niche workflows specific to your repo or your tools, you write the skill yourself or it stays unsupported.

For most general use, the breadth wins. For workflows that are unique to your codebase or your processes, you end up writing skills.

Integrations: the breadth bet

OpenClaw's design decision that most shapes day-to-day use: messaging integrations as first-class. Discord, Telegram, Slack, WhatsApp, and a long tail of others are all built in. The agent process runs on your hardware; the channel connectors plug into the messaging service of your choice.

This is the part that made OpenClaw notable. Before it, most autonomous agent projects assumed a terminal or web UI. OpenClaw made "send the agent a message from your phone, get a result back hours later" feel natural. Anthropic adopting the same shape with Claude Code Channels is the strongest signal that the integration pattern matters.

Model backends and ToS

OpenClaw was originally Claude-centric and has broadened to be model-agnostic. You can point it at:

  • Anthropic's API directly with your own key.
  • OpenAI-compatible endpoints (OpenAI, Azure OpenAI, OpenRouter, etc.).
  • Local open-weight models via Ollama or compatible runtimes.

A specific subtlety worth flagging: Anthropic's ToS adjustments in early 2026 distinguished between "rolling your own agent with Claude OAuth" (not allowed) and "running an agent like OpenClaw with your own Anthropic API key inside an official Claude Code session" (allowed). The Hacker News discussion of those clarifications is the clearest public reference. If you plan to use Claude specifically, check the current ToS state; if you plan to run open-weight models, this constraint does not apply.

The commercial layer

OpenClaw also has commercial offerings that Hermes Agent explicitly does not. By mid-2026 these include enterprise tiers and at least one major partner integration (a NVIDIA NemoClaw offering covered in Cubitrek's comparison).

For users on the open-source side this changes very little — the core project remains MIT-licensed and free. For teams evaluating long-term reliance on a self-hosted agent, the commercial path can be useful (vendor support, hosted variants) or concerning (potential feature gating, two-tier community). Both are real considerations depending on your priorities.

What OpenClaw does not do

Three things worth being explicit about because the project is often described in maximalist terms:

  • It does not autonomously create skills from task history. Skills are written by humans. The agent gets more capable when you (or the community) add skills, not when the agent runs more tasks. If autonomous skill creation is what you want, see How Hermes Agent Actually Works.
  • It is not a polished IDE. OpenClaw does not replace Cursor, Claude Code's editor integration, or similar tools. It works alongside them or independently in messaging channels, but the in-editor experience is not its strength.
  • It is not a managed service. You run the process. You manage the API keys. You decide what the agent has access to. Most failure modes are operational — the model going off the rails, the agent hitting a rate limit mid-task, a skill behaving unexpectedly — and you are the one who has to handle them.

The constraints are part of the design. The reason it works as well as it does is that it picked a coherent set of things to be good at — messaging-first UX, broad skills, model-agnostic — and resisted being everything.

What to study if you are evaluating

If you are choosing between OpenClaw and the alternatives, the materials worth reading directly: