assistant-runtime on tomrochette.com

MetaClaw

tom@tomrochette.com (Tom Rochette) — Fri, 02 Oct 2026 00:00:00 +0000

MetaClaw is an MIT-licensed meta-learning layer that sits in front of OpenClaw-family assistant agents and evolves their skills and cross-session memory from everyday conversations, with optional reinforcement-learning updates scheduled into idle hours.

Its bet is that agent improvement should be a background process the user never operates: talk normally, and a memory layer plus sleep-time training turns each conversation into better skills and recall.

What it is
#

A CLI (metaclaw setup, then metaclaw start) from an academic lab (aiming-lab) with a technical report that ranked first on Hugging Face Daily Papers in March 2026. It wraps multiple assistant harnesses (OpenClaw, IronClaw, PicoClaw, ZeroClaw, CoPaw, NanoClaw, NemoClaw) and runs in three modes: skills-only, reinforcement-learning-only, or auto (skills plus scheduled RL). The RL path trains against Tinker or MinT backends over APIs, so no GPU cluster is required, and slow updates only run during sleep hours, idle time, or calendar meetings, with support/query separation to keep stale rewards out. The memory side persists cross-session context per user and project (facts, preferences, history retrieved into prompts), ingests turns incrementally instead of at session end, and ships as a one-click OpenClaw extension.

Status
#

Dormant-leaning: v0.4.1 released 2026-04-11 and no repository push since 2026-06-07 as of 2026-10-03, roughly four months quiet. The attention record is unusual: first place on Hugging Face Daily Papers and 3,455 stars, but no Hacker News discussion at all, so the buzz came from the paper ranking rather than user adoption. A research artifact with a real feature history through April, then silence; treat it as a promising experiment on pause, not maintained infrastructure.

Strengths
#

The mode menu is genuinely useful: skills-only mode improves behavior with no training stack at all, while auto mode adds scheduled RL for people who want it.
Sleep-hours training with support/query separation addresses the two classic continual-learning failures, catastrophic overwriting and stale reward signals, at the scheduler level.
Incremental memory ingestion (every five turns by default) shrinks the mid-session blackout window most memory layers have.
Multi-claw support means one learning layer covers most of the OpenClaw-family ecosystem, not a single harness.
The paper plus reproducible CLI is a legible research artifact: claims are checkable against the report.

Cautions
#

Four months without a push and no release since April: the OpenClaw family it wraps moves fast, and pinned integration points rot.
RL quality is only as good as the reward signals scraped from conversations, and the note-sized documentation does not include field results beyond the paper’s own benchmarks.
Academic-project risk: the lab’s roadmap, not users, decides whether v0.5 ever happens.
The daily-papers ranking measures paper attention, not deployment safety; nothing here is audited for the failure modes of an agent that rewrites its own skills.
No independent community footprint: no HN threads, no third-party writeups found.

Pricing
#

Free and open source under MIT. Skill-only mode costs nothing beyond your model APIs; RL mode adds the Tinker or MinT training API as a separate bill. No GPU purchase required.

Compared to
#

claude-mem: session-memory compression for Claude Code specifically; MetaClaw is cross-claw and adds training, not just recall.
Cognee: knowledge-graph memory pipelines for applications; MetaClaw closes the loop by feeding learned skills back into the agent.
OpenClaw: the harness MetaClaw most directly extends; use OpenClaw alone for capability, add MetaClaw for automatic improvement.

Bottom line
#

Recommended for OpenClaw-family users and researchers who want a working, cited starting point for self-improving agent memory and skills. Not for production assistants (the project is dormant and unaudited), or anyone who needs a maintained dependency.

Changes
#

2026-10-02 - Created.

References
#

https://github.com/aiming-lab/MetaClaw - repository, MIT license, modes, multi-claw support, memory layer, and the push record as of 2026-10-03
https://arxiv.org/abs/2603.17187 - the technical report “MetaClaw: Just Talk” grounding the meta-learning and RL-scheduling claims
https://huggingface.co/papers/2603.17187 - the Hugging Face Daily Papers ranking (first place, March 2026) behind the attention claim
https://github.com/aiming-lab/MetaClaw/releases - the v0.4.1 release (2026-04-11) anchoring the version and dormancy timeline
https://github.com/aiming-lab/MetaClaw/blob/main/README.md - the mode documentation, incremental-memory changelog, and OpenClaw plugin path

OpenWorker

tom@tomrochette.com (Tom Rochette) — Fri, 02 Oct 2026 00:00:00 +0000

OpenWorker is an MIT-licensed, open-source desktop AI coworker from Andrew Ng’s team that runs assigned tasks to finished deliverables on your own machine, with every action governed and logged and sandboxed execution via NVIDIA’s OpenShell.

Its bet is that the assistant category’s next step is specialist coworkers with governance built in: it ships security-review coworkers first, because attackers already have AI leverage and defenders should too.

What it is
#

A desktop app (macOS Apple Silicon, signed and notarized; Windows x64 builds not yet code-signed) in open beta, self-updating, from the DeepLearning.AI orbit. It runs on your machine, takes a bring-your-own key for OpenAI, Anthropic, Google, or an open-weight provider, or runs fully local through Ollama, and ships specialist coworkers whose tools, working style, and check-ins come preconfigured for one job. The launch cohort is security work: codebase and dependency review where findings combine deterministic scanners (semgrep) with model reasoning and proposed fixes are re-scanned and diff-reviewed before approval, plus cloud-posture audit and incident triage. Every action an agent takes is governed and logged, approvals are configurable, and commands can run inside an NVIDIA OpenShell sandbox.

Status
#

Early and moving fast: about 18,400 stars in its first ten weeks as of 2026-10-03 (created 2026-07-20, pushed 2026-10-02), with v0.3.0 released 2026-09-30 and a public beta disclaimer on the README. The star count is an audience effect as much as an adoption signal: Andrew Ng’s announcement drove the attention, while the independent Hacker News footprint is one 5-point thread plus a pair of 2-point follow-ups, so the tool’s actual field usage is unproven. The project says rough edges are being polished; treat it as beta in fact, not just label.

Strengths
#

Governance is structural, not a settings page: logged actions, approval guidance, and sandboxed command execution are the default path.
The security-coworker design enforces fixer-versus-checker separation, with deterministic scanners anchoring the model’s findings.
No model lock-in: frontier keys or fully local through Ollama, on a machine you control.
Task-oriented framing (finished deliverables, standing automations with full transcripts) rather than another chat window.
Backing and velocity: an Andrew Ng team shipping weekly-quality releases into a category it also teaches.

Cautions
#

Beta software with root-adjacent powers: an agent that reads your code, cloud config, and inbox deserves more scrutiny than its ten-week age suggests.
macOS gets signed builds while Windows builds are unsigned, so the cross-platform story is uneven on the security-sensitive edges.
The coworker roster beyond security is mostly roadmap; the launch promise is narrower than the homepage’s everyday-work examples.
Its governance vocabulary (approvals, logs, sandboxes) overlaps what your OS and IAM already enforce, and the two layers are not integrated.
The 18k-star traction has no independent usage record behind it yet.

Pricing
#

Free and open source under MIT. You pay your model provider directly, or nothing if you run Ollama locally; there is no hosted tier and no OpenWorker subscription.

Compared to
#

OpenClaw: the heavyweight personal assistant framework with channels, skills, and a huge ecosystem; choose OpenClaw for maximal capability, OpenWorker for governed, task-finished security work from day one.
Hermes Agent: Nous Research’s autonomous agent with self-created skills and messaging gateways; OpenWorker is the desktop-first, governance-first counterpart.
OpenShell: the NVIDIA sandbox OpenWorker builds on; use it directly if you want the isolation without the coworker.

Bottom line
#

Recommended for security-minded engineers who want a governed desktop coworker for review and triage work they can run on their own keys. Not for anyone needing a mature, broadly capable assistant today, or Windows-first teams that need signed builds.

Changes
#

2026-10-02 - Created.
2026-10-03 - Corrected the Hacker News footprint claim to one 5-point thread plus a pair of 2-point follow-ups (the Algolia search returned a third thread the creation run missed), and refreshed the as-of numbers (about 18,400 stars, pushed 2026-10-02, v0.3.0 still latest).

References
#

https://github.com/andrewyng/openworker - repository, MIT license, stars, beta disclaimer, use cases, and the governed-by-design section as of 2026-10-03
https://openworker.com/ - product positioning, download platforms, and the specialist-coworker framing
https://github.com/andrewyng/openworker/releases - the v0.3.0 release (2026-09-30, still latest as of 2026-10-03) anchoring the version claim
https://github.com/andrewyng/openworker/blob/main/docs/openshell.md - the NVIDIA OpenShell sandbox integration behind the isolation claims
https://github.com/andrewyng/openworker/blob/main/docs/approval-guidance.md - the approval model behind the governance claims

assistant-runtime on tomrochette.com

MetaClaw

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

OpenWorker

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

assistant-runtime on tomrochette.com

MetaClaw

What it is #

Status #

Strengths #

Cautions #

Pricing #

Compared to #

Bottom line #

Changes #

See also #

References #

OpenWorker

What it is #

Status #

Strengths #

Cautions #

Pricing #

Compared to #

Bottom line #

Changes #

See also #

References #

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#