meta-harness on tomrochette.com

Omnigent

tom@tomrochette.com (Tom Rochette) — Tue, 06 Oct 2026 00:00:00 +0000

Omnigent is Databricks’ open-source meta-harness: an orchestration layer that wraps Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and your own YAML-defined agents behind one uniform API, then adds policies, sandboxing, and real-time collaboration on top.

What it is
#

It sits above the harness, not beside it: sessions, policies, and history live in Omnigent, while the model-carrying harness underneath stays swappable with one-line changes. A runner wraps any agent in a sandboxed session with a uniform API, a local or deployed server provides policies and shared history, and every session is reachable from the terminal, the web, mobile, or the macOS desktop app. Policies stack across server, agent, and session levels to cap spend, pause risky actions for approval, and broker credentials through an egress proxy so the agent never sees the token. Agents are short YAML files (prompt, harness, tools, sub-agents); the same file runs on any supported harness, and example agents include Polly, a coding orchestrator that delegates to parallel worktree sub-agents and routes each diff to a reviewer from a different vendor. Apache-2.0, Python 3.12+, installed via installer, uv, pip, or Homebrew; built by the Databricks AI team with Neon and contributors, announced June 13, 2026 by Matei Zaharia, Kasey Uhlenhuth, and Corey Zumar, in alpha.

Status
#

Active, well-resourced, and young, with traction that far outruns its independent discussion. 10,601 stars and 1,695 forks as of 2026-10-06 (repository created 2026-06-11), PyPI 0.17.0 published 2026-10-06 after 40+ releases, and the default branch was pushed the same day. The HN footprint is thin: the top thread (the Databricks announcement, June 2026) reached 15 points with 4 comments, and resubmits stayed at 1 to 6 points. The 1,814 open issues against 10.6k stars read as heavy use plus heavy traffic, under an alpha badge the README still wears.

Strengths
#

Harness portability as the product: the same agent YAML runs on claude, codex, cursor, hermes, pi, and opencode harnesses plus SDK harnesses, with a harness test bench that checks each one’s capability matrix against observed behavior.
Governance at the right layer: stateful contextual policies (cost budgets, tool restrictions, approval gates) enforced outside the prompt, where the agent cannot argue with them.
Sandboxing with teeth: bubblewrap on Linux (mandatory), seatbelt on macOS, and cloud sandboxes on Modal, Daytona, Blaxel, E2B, Kubernetes, Databricks, and more, includable per session.
Collaboration: live session sharing by URL, co-driving, forking, and the same session from terminal, web, phone, or desktop app.

Cautions
#

Alpha at 0.x: expect churn (40+ PyPI releases in under four months), and the README says so itself.
It is another layer to own: a local server at minimum, optionally a deployed one, plus credential routing and sandbox providers; if you run one harness happily, this is overhead, and an HN commenter made exactly that trade against its sandbox abstraction.
Windows support is explicitly degraded: no OS sandbox, no terminal wrappers, Job Object containment only.
Most evidence is the vendor’s own: the thin independent discussion means no second opinion on real-world governance behavior yet.

Pricing
#

Apache-2.0 open source; no paid tiers surfaced as of 2026-10-06 (the site and README document no plans; the costs are your model tokens and any cloud sandbox hosting you attach), so pricing does not apply.

Compared to
#

Pi: the minimal, extension-first single harness; choose Pi to go deep on one harness, Omnigent to rotate harnesses without rewriting agents.
OpenCode: the provider-neutral single harness; Omnigent treats it as one more engine in the rotation.
Warp Agent CLI: the commercial, credit-metered agent platform; Omnigent is the open governance layer over the harnesses you already run.

Bottom line
#

Recommended for teams running several coding harnesses that need shared policy, sandboxing, and session collaboration above them, and who accept alpha churn. Not for developers happy inside one harness, and not for Windows-first setups that need OS-level isolation.

Changes
#

2026-10-06 - Created.

References
#

https://api.github.com/repos/omnigent-ai/omnigent - repository facts (10,601 stars, 1,695 forks, Apache-2.0, Python, pushed 2026-10-06, 1,814 open issues, as of 2026-10-06)
https://raw.githubusercontent.com/omnigent-ai/omnigent/main/README.md - harness list, agent YAML, policies, sandboxes, install, Windows limits, collaboration
https://www.databricks.com/blog/introducing-omnigent-meta-harness-combine-control-and-share-your-agents - Databricks origin (Zaharia, Uhlenhuth, Zumar, June 13, 2026), architecture, Apache-2.0 open-sourcing
https://omnigent.ai - product site (alpha status, “Built by the Databricks AI team, Neon and the Omnigent Contributors”)
https://hn.algolia.com/api/v1/search?query=omnigent - the thin HN footprint (15-point top thread of June 2026; resubmits at 1 to 6 points)
https://pypi.org/pypi/omnigent/json - PyPI 0.17.0 published 2026-10-06, Python 3.12+, release history

Ruflo

tom@tomrochette.com (Tom Rochette) — Tue, 06 Oct 2026 00:00:00 +0000

Ruflo is an MIT-licensed, npm-installed meta-harness that wraps Claude Code and Codex with 100+ prebuilt specialized agents, coordinated swarms, persistent self-learning memory, and cross-machine federation.

Ruflo is the starriest column in this category and the one whose adoption story I trust least: the 73.9k stars measure the renamed Claude Flow repository, the download badges link to proof files committed inside the project’s own repo, and the entire independent footprint I can find is a three-point Show HN thread.

What it is
#

A layer around Claude Code and Codex from rUv Cohen’s Cognitum.One: one npx ruflo init installs 100+ agents, swarm coordination, a learning loop, vector RAG, and security guardrails on top of the CLIs you already run. The README’s own framing is “Agent = Model + Harness”, with Ruflo as the harness: an execution layer that routes work from CLI and MCP surfaces through a router into swarms, memory, and model providers. Federation lets installations on different machines talk to each other without leaking data, a claim no other column in this category’s matrix makes at the agent level. It ships under MIT, installs from npm, and drives the Claude Code and Codex accounts you already have.

Status
#

Active and shipping almost daily: v3.52.1 released 2026-10-06, v3.52.0 the day before, with 40 contributors and an npm release feed that still carries claude-flow-named packages. 73,943 stars, 8,795 forks, and 1,113 open issues as of 2026-10-06. npm ruflo served 60,190 downloads in the week ending 2026-10-04. Those stars are inherited: github.com/ruvnet/claude-flow returns a 301 redirect to ruvnet/ruflo, and the README’s star badge still points at the claude-flow shields URL, so the number measures the repository’s history under the Claude Flow name plus a rename, not adoption of Ruflo the product. The independent evidence for that product is thin: a Show HN from July 2026 at 3 points and an April 2026 “Is any one using ruflo?” post at 1 point. The “8.1M+ ecosystem downloads” and “106k git clones in 14 days” badges link to JSON proof files committed inside the repository, which is self-measurement, not third-party measurement.

Strengths
#

Breadth per install: agents, swarms, memory, RAG, hooks, and federation in one npx ruflo init, where most columns here do one of those.
Shipping cadence: multiple releases in the first days of October 2026, and 60k weekly npm installs even after discounting the self-reported badges.
Cross-machine federation is a capability claim no other column in the matrix makes at the agent level.
MIT and provider-agnostic at the model layer: it drives the Claude Code and Codex subscriptions you already pay for.

Cautions
#

The provenance problem: the project built its reputation as Claude Flow, renamed to Ruflo, and kept the number, while the README badge still points at the old URL and old write-ups describe a different scope.
Every adoption number that matters is self-reported: the badges cite in-repo proof JSONs, the fourteen-chapter guide is written by the founder, and it is a republished LinkedIn post.
The community footprint contradicts the star count: 3 points and 1 point on Hacker News against 73.9k stars is the widest gap I have measured in this category.
1,113 open issues as of 2026-10-06 suggests support load that has outrun the maintainers.
The scope is enormous (memory, RAG, swarms, federation, guardrails) for a project past major version 52, which is churn risk by definition.

Pricing
#

Free and open source under MIT. No hosted tier and no Ruflo bill: agents run on your existing Claude Code and Codex accounts, so inference is billed by those providers.

Compared to
#

OpenRig: the other local meta-harness over your existing Claude Code and Codex installs; choose OpenRig for named, snapshot-restorable teams of terminal sessions, Ruflo for in-process swarms with memory and federation.
oh-my-codex: the workflow layer for Codex CLI alone; choose it if Codex is your only CLI, Ruflo if you run Claude Code and Codex and want agents, memory, and federation around both.
CrewAI: the Python framework where you assemble the crew yourself; choose CrewAI when you must own the orchestration code, Ruflo when you want the crew prebuilt.

Bottom line
#

Recommended for Claude Code and Codex users who want a maximal prebuilt agent layer and will verify claims against their own usage rather than the repository’s badges. Not for teams that need an independently verified adoption history, a stable API surface, or a scope that fits in one afternoon.

Changes
#

2026-10-06 - Created.

References
#

https://api.github.com/repos/ruvnet/ruflo - stars, forks, open issues, MIT license, and push date as of 2026-10-06
https://raw.githubusercontent.com/ruvnet/ruflo/main/README.md - the meta-harness definition, the 100+ agents claim, the claude-flow badge URLs, and the in-repo proof badges
https://api.github.com/repos/ruvnet/ruflo/releases - v3.52.1 (2026-10-06), the daily cadence, and the claude-flow-named npm packages
https://raw.githubusercontent.com/ruvnet/ruflo/main/docs/ruflo-explained.md - the fourteen-chapter guide and its Cognitum.One authorship
https://api.npmjs.org/downloads/point/last-week/ruflo - 60,190 downloads in the week ending 2026-10-04
https://github.com/ruvnet/claude-flow - the 301 redirect to ruvnet/ruflo behind the inherited star count
https://cognitum.one/agentic-engineering - the company page the README banner promotes
https://hn.algolia.com/api/v1/search?query=ruflo&tags=story - the thin Hacker News footprint, both threads and their point counts
https://news.ycombinator.com/item?id=49051072 - the Show HN thread at 3 points, July 2026
https://news.ycombinator.com/item?id=47943679 - “Is any one using ruflo?” at 1 point, April 2026

meta-harness on tomrochette.com

Omnigent

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

Ruflo

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

meta-harness on tomrochette.com

Omnigent

What it is #

Status #

Strengths #

Cautions #

Pricing #

Compared to #

Bottom line #

Changes #

See also #

References #

Ruflo

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
#