Steinberger’s decision to join OpenAI triggered the single biggest structural change in OpenClaw’s eight-year history: the project is leaving Peter’s one-person GmbH and moving into an independent foundation. If you maintain an OpenClaw bot, run ClawCloud in production, or just like reading governance docs for fun, this post collects everything that’s public so far, adds context from similar foundations, and bluntly lists the risks.

Why the switch to a foundation at all?

OpenClaw started as clawdbot in 2016. For years Peter signed the CLA, paid AWS bills, and cut npm releases himself. That worked while the project had <10 contributors and the legal exposure fit on a napkin. Today:

  • GitHub stars: 145,000+
  • Weekly npm downloads: >1.3 M
  • 750+ paying ClawCloud orgs (last public number, Q1-2024)
  • 900+ third-party plugins in the registry
  • 1 critical CVE every ~7 months (average of the last three years)

That scale demands liability protection, predictable IP ownership, and a public roadmap that doesn’t depend on one person’s calendar. The foundation gives each of those a dedicated organ instead of hoping Peter stays healthy and online.

Foundation paperwork: what’s on the table

The draft charter lives in org/openclaw-foundation/rfcs/0001-foundation.md (commit f3a2e7d, May-29). The filing jurisdiction is Austria because that’s where PSPDFKit GmbH already has the accountants on speed dial. Legally, we’re looking at a gemeinnützige Privatstiftung—the Austrian equivalent of a U.S. 501(c)(6) trade association.

Key provisions in the draft:

  • Copyright assignment: Contributors sign a CLA that grants the foundation a broad, irrevocable license. No copyright transfer. GPL + MIT dual-license stays.
  • Asset transfer: GitHub org, npm namespace, and the openclaw.dev domain move to the foundation on day 0.
  • Trademark: The foundation owns the OpenClaw mark and enforces usage. This fixes the 2022 Clawdbot trademark mess that led to the rename.
  • Fiscal year: July 1 – June 30 to sync with OpenAI’s FY and simplify grant reporting.

Missing bits: the final bank account (they’re still picking between Erste Bank and N26 Business) and the insurance carrier for D&O coverage. Both are required before the notary signs off, so the official launch date—currently penciled in for September 18, 2024—could slip.

Governance model: board, TSC, and voting mechanics

The foundation borrows heavily from the CNCF playbook but with fewer layers:

Board of Directors (BoD)

  • 5 seats, 2-year terms, term-limit: 3 consecutive terms
  • 1 seat reserved for the “Lead Sponsor” (that’s OpenAI for the first cycle)
  • 1 seat for “Community Maintainers” (elected by committers with >200 merged PRs)
  • 3 at-large seats elected by paid foundation members (companies >€5k/yr)

The board handles money, trademarks, and can hire/fire staff. It cannot merge code or set technical direction.

Technical Steering Committee (TSC)

  • 7 seats, 1-year terms, no term limits
  • Seats allocated by the total number of merged PRs in the last 24 months—pure meritocracy, no corporate overrides
  • Has final say on the roadmap, release cadence, and technical RFCs
  • Can veto board decisions that “materially affect the technical direction” with a 5/7 super-majority

In practice, that means the board can’t, for instance, force a relicensing to Apache-2.0 if the TSC blocks it. But the board can decide to spend €250k on a security audit without TSC approval.

OpenAI sponsorship: confirmed numbers and strings attached

OpenAI’s public blog post (May-31) committed to three things:

  1. €1 M in unrestricted funding over 24 months, paid quarterly
  2. $400k in Azure compute credits earmarked for CI runners and canary environments
  3. “Priority access” to ChatGPT model previews for OpenClaw bug-bash events

The money is unrestricted; the board controls allocation. However, there’s a claw-back clause: if the foundation ever adopts a license that prevents commercial use, the remaining funds revert to OpenAI. That’s standard fare, but still worth noting for the “pure copyleft or die” crowd.

No seat on the TSC is reserved for OpenAI employees (Steinberger explicitly refused that). The only formal influence is the single BoD seat mentioned earlier.

Community input: how the rest of us get a voice

Three channels land in the bylaws:

  1. Open RFC calls. Every second Wednesday, 16:00 UTC on Zoom. Anyone can propose an agenda item by opening issues/rfc/<date>.md.
  2. Annual contributor summit. First one planned for March 2025 in Vienna. Travel grants budgeted at €30k.
  3. Binding votes. If 100+ unique contributors sign a GitHub discussion with 👍 within 14 days, the TSC must add the issue to its next meeting agenda and publish minutes.

Is that enough? Maybe. Kubernetes governance took five years to evolve something similar. At least it’s explicit from day 1.

Roadmap implications: what actually ships and when

Here’s the headline: nothing ships on time if governance churn eats the next three months. Assuming the paperwork closes in September, the draft roadmap looks like this (pulled from docs/roadmap/2024-2026.md):

  • Q4 2024: Gateway v5 with WebAssembly plugin sandboxing (was v4.4 in old roadmap)
  • Q1 2025: Memory subsystem rewrite to SQLite 4 with WAL checkpointing (kills the dreaded --max-mem flag)
  • H1 2025: Initial Rust daemon prototype behind a --experimental-rs flag; aims for 5× lower idle CPU use
  • H2 2025: Scheduler API GA; lets agents register cron-like tasks in 2 loc
  • 2026: “Composable Agents” spec v1 submitted to the W3C community group Steinberger teased on Slack

A few things to unpack:

Gateway v5 and WebAssembly sandboxing

Right now, OpenClaw plugins are Node ESM modules loaded in-process. That’s a vector for require('fs') shenanigans. The WASM rewrite isolates plugins to a bytecode boundary. The catch: Node 22’s WASI support is still flagged --experimental-wasi-unstable-preview1. Expect rough edges.

SQLite 4 migration

The persistent memory subsystem currently stores JSON blobs in LevelDB. Large deployments either shard with Redis or give up and disable memory. The SQLite 4 rewrite uses the new “imperfect indexes” feature to keep vector embeddings and scalar metadata in the same file. Early benchmarks: 4.2× faster recall, 37 % smaller disk usage.

Rust daemon prototype

Peter has said for years that Node’s event loop is “fine until it isn’t.” The daemon owns long-running shell tasks and browser automation, so memory leaks hurt. The Rust rewrite lives in openclaw-rs, currently 2 weeks old. Don’t cargo install it yet unless you like double-digit segfaults.

Speculation: Steinberger’s fingerprints at OpenAI and what that means

The elephant in the room: Peter is now Senior Director of Agents at OpenAI. Conflict? Opportunity? Both. A few educated guesses:

  • Model integrations will skew OpenAI first. Anthropic support landed in 0.33.0, Gemini in 0.35. OpenAI models will still be the default, likely with deeper function-call hooks.
  • SAML and SOC-2 work will accelerate. OpenAI’s enterprise sales team needs compliant glue for customers; expect ClawCloud to inherit those checkboxes.
  • License stays permissive. OpenAI won’t sink money into a copyleft project. MIT/GPL dual is the compromise; GPL satisfies OSS purists while MIT keeps the commercial doors open.
  • OpenAI may upstream tooling. Rumor: their internal agent evaluator (codename “Kraken”) might be donated. If true, expect a wave of PRs adding evaluation harnesses.

Could OpenAI just acquire the whole thing later? With the foundation in place, it’s legally harder but not impossible. The assets live in the foundation, so an acquisition would require a super-majority board vote and, because of the veto power, probably TSC consent. Not cheap, not quick.

Risks and trade-offs worth calling out

1. Bureaucracy overhead. Five board seats mean five calendars. Releases slow when governance meetings drag on. The Node.js project spent 18 months fighting over similar structures.

2. Single sponsor dependency. €1 M is great until it isn’t. If OpenAI business priorities shift, renewal might evaporate in 2026. The foundation needs at least two more “lead sponsors” to avoid vendor lock-in.

3. Talent drain. Peter isn’t the only maintainer fielding job offers. If the TSC loses two senior maintainers, the Rust rewrite stalls.

4. License flame wars. Someone will propose moving to BSL or Elastic-2.0 the minute a cloud vendor launches a paid SaaS. The bylaws allow it with enough votes, so brace for drama.

How to get involved today

If you care about where OpenClaw goes after the hand-off, you have three concrete moves:

  1. Review the charter PR #42 and leave line comments before June 30.
  2. Sign the interim CLA: curl -X POST https://cla.openclaw.dev/sign -d "github=<username>"
  3. Test the foundation branch (Node 22+ required):
    npm create openclaw@canary -- --gateway v5 --memory sqlite4 cd my-agent OPENCLAW_EXPERIMENTAL_WASI=1 npm start

The foundation launch is paperwork, but the code still needs reviewers. If governance does its job, the next year should bring cleaner plugins, fewer segfaults, and a saner release process. If it doesn’t, we’ll be back here complaining. Your call.