Assistant Runtimes Feature Matrix

This matrix compares the nine assistant runtimes profiled in this section: the OpenClaw root, the three variants named after shrinking it, the learning-loop challenger, the readable Python core, the two Cowork-style desktops, and the channel-matrix newcomer. Everything below was verified against live sources on 2026-08-30 and re-verified on 2026-09-02, 2026-09-04, 2026-09-05, 2026-09-06, 2026-09-07, 2026-09-08, 2026-09-09, 2026-09-10, 2026-09-12, and 2026-09-13.

The family ladder is a trust ladder: OpenClaw is an ecosystem, NanoClaw an auditable codebase, ZeroClaw a static binary, PicoClaw a firmware image, Hermes a memory that grows, and the 2026 columns stretch the ladder again, Nanobot bets on readable Python, QwenPaw on channels and security defaults, and OpenWork and Eigent carry the category onto the desktop as Cowork alternatives.

Legend: ✓ supported, ✗ not supported, ~ partial or conditional, ? not verified as of the date above. Each column links to the full research note; every cell below traces to a source cited there or in the references.

The matrix

#

Reading the matrix

#

The isolation and credentials rows are the ones that bite: only NanoClaw containers the agent and vaults the keys by default, QwenPaw ships the strongest default posture (five security layers including a kernel sandbox), while OpenWork and Eigent run on your host by design, the trade for being desktop apps. PicoClaw’s isolation is a recent feature flag rather than an architecture, and ZeroClaw’s answer is that the binary itself is small enough to trust.

The providers row hides the family’s politics: the root endured Google and Anthropic restricting subscriptions for it in 2026, so check your provider’s current stance toward this family before standardizing. QwenPaw is the only column with a real offline path through its own trained small models.

The desktop split is the new category line: OpenWork and Eigent do not do chat channels at all, which is why their cells go tilde there, and choosing between them and the messaging runtimes is really choosing between an assistant that lives in your chats and one that lives on your desktop.

The status row’s issue counts are inversely proportional to age, not quality: the root carries 7,078 open issues at scale, and the pre-1.0 PicoClaw carries 51, so read the column against its birthday.

Hermes is the column that breaks the -claw pattern: it competes on the learning loop rather than the trust ladder, and its about 245k stars against a 42.4k-issue backlog is the trade in one row pair.

Choosing from the matrix

#

• Want the ecosystem, channels, and companion apps and accept the weight: OpenClaw.
• Want the assistant to compound (skills, memory, user model) and will supervise it: Hermes.
• Want to read every line and cage the agent: NanoClaw.
• Want to stop needing to read it, or run on a Pi: ZeroClaw.
• Want it on a microcontroller or salvage hardware, pre-1.0 accepted: PicoClaw.
• Want a readable Python core you can extend: Nanobot, binding to localhost and owning the security surface.
• Want the deepest chat-channel matrix including Chinese apps, or a real offline assistant: QwenPaw.
• Want a Cowork-style desktop on OpenCode with skills that port via MCP: OpenWork.
• Want a Cowork-style desktop with multi-agent workforces under Apache-2.0: Eigent.

Changes

#

• 2026-08-27 - Created with four columns and twelve rows and the trust-ladder thesis.
• 2026-08-27 - Extended to five columns with Hermes inserted, and the reading section gained the learning-loop-versus-trust-ladder line.
• 2026-08-30 - Nanobot, OpenWork, Eigent, and QwenPaw columns added, matrix at nine columns, with desktop-split prose added.
• 2026-09-13 - Removed the archived Android port from ZeroClaw’s edge and mobile cell and corrected its open-issue count to 824.

See also

#

• Control Planes Feature Matrix - the layer that manages these as employees
• Harness Feature Matrix - the coding-agent cousins
• Read the Commits, Not the Manual - maintaining the root at scale
• Managing Many Concurrent LLM Agent Sessions - the supervision context

References

#

• https://github.com/openclaw/openclaw - Gateway model, channels, security posture
• https://github.com/NousResearch/hermes-agent - learning loop, backends, providers for the Hermes column
• https://github.com/nanocoai/nanoclaw - container isolation, vault, skill-based channels
• https://github.com/zeroclaw-labs/zeroclaw - runtime model, provider and channel counts
• https://github.com/sipeed/picoclaw - footprint, architectures, security banners
• https://github.com/HKUDS/nanobot - stars, alpha status, and surfaces for the Nanobot column
• https://github.com/different-ai/openwork - the OpenWork column: OpenCode base, split license, adoption
• https://github.com/eigent-ai/eigent - the Eigent column: CAMEL workforce, Apache-2.0, maturity
• https://github.com/agentscope-ai/QwenPaw - the QwenPaw column: channels, security layers, offline models