Peter Steinberger’s 64,048 GitHub Commit Contributions: The Machine Behind OpenClaw

Twelve months ago, OpenClaw did not exist.

Peter Steinberger was already building at an unusual pace. He had returned from a long post-PSPDFKit break, spent the first half of 2025 learning how far coding agents could be pushed, and built VibeTunnel as a browser-accessible terminal controller for steering those agents. But on 23 July 2025 there was no Clawd, no Moltbot, no OpenClaw Foundation and no repository receiving tens of thousands of pull requests.

One year later, GitHub attributes 64,048 public commit contributions and 6,855 pull-request contributions to Steinberger’s @steipete account across the measurement window. Its collection-level fields report commits in 169 repositories and opened pull requests in 147. OpenClaw accounts for 37,573 of the commit contributions, but the other 26,475 are the more revealing part of the story: CodexBar, Peekaboo, Oracle, Summarize, gogcli, ClawSweeper, Crabbox, RepoBar, Poltergeist and a long tail of agent-facing tools.

The API result is real. Reading it as 64,048 hand-written commits is not.

Some attributed changes were generated by agents. Some were small. Many repositories were experiments rather than products, and a large community produced much of the work visible in the OpenClaw organisation. Commit volume measures activity, not quality.

What it measures is something more interesting: the event trail left when an experienced product engineer restructures software development around several coding agents, gives those agents tools that close their own feedback loops, and then applies that system to an open-source project under extreme community demand.

This is the story of how that machine formed—and where it broke.

The 12-month scorecard

We queried GitHub’s GraphQL contribution collection for @steipete, inspected repository metadata, paginated the OpenClaw release and contributor endpoints, and cross-checked pull requests through GitHub Search. The query runs from 23 July 2025 to 23 July 2026. GitHub’s date-bucketed response included the partial 23 July contribution day visible when we retrieved it at 07:19 UTC.

The 169 and 147 figures are collection-level totals. GitHub exposes at most 100 named commit repositories and 100 named PR repositories; those two inspectable lists contain 117 distinct names between them. The totals tell us the footprint was larger than the repository list we can audit individually. GitHub also displayed 198,789 restricted contributions without repository or activity detail; those remain excluded from the named-repository analysis.

The four completed phases through 22 July contain 64,000 commit contributions and 6,807 PR contributions: 175.3 and 18.6 per day across 365 completed days. The partial 23 July bucket lifts the headline totals to 64,048 and 6,855. Even those averages hide the shape of the year.

The four phases exclude the partial final day. Their boundaries were drawn so no calendar date appears in two adjacent windows.

The first lesson is visible immediately. The decisive jump happened before OpenClaw became a global project. The second jump was different: not just more commits, but an enormous increase in pull-request activity. The workflow first became a tool factory, then a community integration machine.

Before the machine: losing and finding the reason to build

Steinberger’s output cannot be explained by a model release alone.

He had already spent 13 years building PSPDFKit, a developer-tools company that grew far beyond a solo codebase. In his June 2025 essay “Finding My Spark Again”, he described leaving after selling his shares and feeling that the company had consumed his identity. He stopped building for years. The return came when creating software became exciting again, and AI made the gap between an idea and a working system feel radically smaller.

That history supplied two scarce inputs that commit statistics cannot show.

The first was product judgement: deciding what should exist, what should feel delightful, and which inconvenience is worth turning into software. The second was management judgement. In his February 2026 conversation with Lex Fridman, Steinberger connected his comfort with imperfect agent output to having led engineering teams. Employees did not write code exactly as he would; progress required giving them room, preserving the architectural picture and intervening where consequences mattered.

Agents changed the cost of execution. They did not supply the accumulated taste that selected and steered the work.

That distinction is the foundation of the whole year. Steinberger stopped treating code as the product of his fingers and started treating it as the output of a system he managed.

July to October: from one assistant to a terminal full of workers

At the beginning of the window, that system was still changing quickly.

In “My Current AI Dev Workflow”, published on 25 August, Steinberger described Claude Code as his primary driver. He generally ran one or two agents, increasing to roughly four for cleanup, tests or interface work. He used GPT-5 to review larger plans, kept the agents on main, and preferred ordinary command-line tools over elaborate integrations.

The important design rule was already present: close the loop.

An agent that can only write code creates a human verification queue. An agent that can compile, lint, run tests, inspect logs, query a database and exercise a UI can detect and repair many of its own mistakes. Steinberger increasingly selected tools and architectures on that basis. A CLI was not merely a convenient interface for him; it was an executable interface for the model.

By 14 October, “Just Talk To It” documented a different operating mode. Codex had become the daily driver. Steinberger said agentic engineering now wrote “pretty much 100%” of his code. He ran three to eight Codex sessions in parallel, often in the same repository, while the agents made atomic commits themselves.

The human role moved up a level:

• estimate the blast radius of each change;
• keep simultaneous tasks from colliding;
• discuss architecture and dependencies;
• stop an agent when runtime suggested it had misunderstood the system;
• provide screenshots or existing implementations as compressed context;
• ask for tests while the implementation context was still loaded;
• refactor continuously rather than postponing cleanup.

This was not autonomous software creation. Steinberger remained the scheduler, product manager, architect and escalation point. Parallelism increased execution capacity, but it also increased the number of mental models he had to hold.

The 1,472 commit contributions in the first phase are substantial. The 17,270 in the next phase show when the method began compounding.

November and December: the tool factory reaches critical mass

Repository creation makes the transition visible. Steinberger created two original repositories under his account in September and one in October. Then came 11 in November and 19 in December.

Release activity shows that this was not only repository churn. Across non-fork repositories owned by @steipete, 37 shipped at least one public GitHub release during the 12-month window. CodexBar alone published 91 releases.

Not all 30 were durable products. Some were experiments, support repositories, websites or narrow utilities. That is precisely the point. When implementation becomes cheap, the threshold for turning friction into software collapses.

Several November projects became structural parts of the workflow:

• agent-scripts, created 8 November, shared instructions and scripts across repositories;
• oracle, created 14 November, let an agent package files and a hard question for GPT-5 Pro;
• CodexBar, created 16 November, exposed coding-agent usage and limits;
• RepoBar, created 24 November, surfaced repository health, CI, issues, PRs and releases.

December filled in the world around the future personal agent. There were tools for cameras, voice, Sonos, Eight Sleep, browser cookies, phone calls, food delivery, summarisation and terminal-friendly rendering. Each solved a small problem. Together they made more of Steinberger’s environment legible and controllable through commands.

The timeline suggests an underappreciated link to OpenClaw. The project did not spring from one perfect specification. It emerged after months spent converting services, devices and workflows into things an agent could call and verify.

In “Shipping at Inference-Speed”, published 28 December, Steinberger wrote that output was becoming limited mostly by “inference time and hard thinking.” His normal pattern was one large project surrounded by several satellite projects. He reused solutions across repositories by pointing Codex at an existing sibling project, maintained agent-readable docs, queued follow-up tasks and designed codebases for models to navigate.

The factory was not one model. It was a feedback system:

1. encounter friction;
2. turn it into a CLI, library or instruction;
3. let agents use the new tool;
4. reuse the pattern across other repositories;
5. spend the recovered time removing the next constraint.

OpenClaw became the integration point for that accumulated infrastructure.

24 November: prompting the missing assistant into existence

The repository now known as openclaw/openclaw was created on 24 November 2025. Its early history used the name warelay; the first public release, v0.1.1, followed under that name on 25 November.

The initial idea was narrower than the present platform: connect a command-line agent to WhatsApp. Steinberger told Fridman that he had wanted a personal assistant since April but assumed the major labs would build one. By November he was annoyed that it still did not exist, so he “prompted it into existence.”

He said the first prototype took roughly an hour. The important moment came later, while Steinberger was travelling in Morocco. As he later told Fridman, he sent the agent a voice message even though the relay only explicitly handled images. The agent identified the audio format, converted it, found an available transcription path and replied. The code had not been written specifically for that sequence; the model used tools already present in the environment.

In Steinberger’s telling, that incident widened the product hypothesis. A capable model with broad but comprehensible access to a computer did not need a bespoke feature for every action. It needed tools, context, permissions and a channel through which the user could ask.

The repository evolved rapidly from warelay into a sequence of public and package identities. Peter called the project Clawd publicly. Release v2.0.0-beta1 rebranded the package to Clawdis on 19 December while adding the macOS companion, WebSocket Gateway and iOS node; the 5 January release line was named clawdbot. By then the recognisable Gateway architecture was in place: one process owning channel and session state, with command-line, web and app surfaces connecting as clients. The satellite tools—Peekaboo for computer control, gogcli for Google services, mcporter for on-demand MCP access, voice and home-control utilities—became the limbs of the larger system.

On 29 January 2026, the official OpenClaw announcement documented the final naming sequence. It said Anthropic had contacted the project about the similarity between Claude and Clawd. Moltbot followed briefly. OpenClaw became the durable name after trademark checks, domain purchases and migration work.

Two months after the first prototype, the announcement reported more than 100,000 GitHub stars and said the project had already outgrown solo maintenance.

Where GitHub attributed the 64,048 commit contributions

OpenClaw dominates the account, but it does not erase the supporting system.

The top five repositories contain 73.82% of the commit contributions; the top ten contain 81.94%. That concentration argues against a picture of random repo spam. The long tail was large, but most credited activity flowed into a small set of repeatedly used systems.

It also shows why “Peter built OpenClaw” is incomplete. OpenClaw was the centre of gravity, while a constellation of tooling repositories made the centre productive.

January to April: from solo velocity to community pressure

OpenClaw’s viral growth changed the work.

In his February interview with Fridman, Steinberger described pull requests as “prompt requests”: external contributions were often most valuable as evidence of user intent. He would ask Codex to identify the underlying problem, inspect the wider architecture and decide whether the submitted implementation was the best solution. Small PRs could take longer to review safely than recreating the fix from a clear issue.

Whatever contributors make of the terminology, the operational distinction is real. A pull request from a stranger introduces a different trust boundary from agent-generated work based on the maintainer’s own prompt. Steinberger told Fridman that he read more external code for precisely that reason. He wanted the problem, context and prompt—not only a patch detached from intent.

GitHub’s attribution supports the scale of this mode. Through 22 July, public search found 4,629 OpenClaw PRs authored by @steipete, of which 4,269 were merged. GitHub’s contribution graph records 4,685 OpenClaw PR contributions. That 56-PR gap combines a window mismatch—the graph includes the partial 23 July bucket—with GitHub’s different attribution and indexing rules.

Across all public repositories, Search found 6,795 authored PRs and 6,357 merged. That 93.55% merged share does not show that every first attempt was accepted. It shows an account using PRs as a high-frequency integration and delivery mechanism.

The repository around him was much larger still. At retrieval, OpenClaw had:

• 71,481 commits reachable from its default branch;
• 65,864 pull requests created;
• 18,533 pull requests merged;
• 2,951 contributors returned by the paginated contributors endpoint;
• 226 GitHub releases, including 106 stable releases and 120 prereleases.

Those are project totals, not Peter totals. They include a global community and should never be added to his account activity. They matter because they describe the input pressure applied to his workflow.

On 15 February, Steinberger announced that he was joining OpenAI to work on bringing agents to a wider audience. He said OpenClaw would move into a foundation, remain open and independent, and continue supporting multiple models and companies. His stated reason was not an exit from the project: he wanted access to frontier research and a route to make an agent that even his mother could use.

April and May: velocity sends a bill

A productivity profile that ends at the commit graph would miss the most important counterexample.

In late April, OpenClaw shipped regressions around plugin dependency repair, startup and update paths, channel behaviour and Gateway performance. Users downgraded. On 5 May, Steinberger published “OpenClaw Had a Rough Week”, apologised, and said the project had underestimated the difficulty of separating optional plugins from core.

He also named the organisational failure: OpenClaw was still too founder-driven. Release, review, packaging and support work remained concentrated around him.

This matters because the final quarter was also his most active: 30,295 commit contributions and 6,513 PR contributions through 22 July. Activity more than doubled from the previous phase, but the operational lesson was not “ship even faster.” It was that a project used as infrastructure needs boring release discipline, stable ownership and a team capable of absorbing review and support.

Commit volume tells us that change happened. It cannot tell us that the change was correct.

The same caveat applies to agent-written code. Steinberger has consistently said he does not read every boring transformation or interface detail, not that no code deserves review. He has identified database paths, security boundaries and untrusted external contributions as areas requiring greater scrutiny. His workflow pushes verification into tests, local execution, linters, model review and iterative refactoring. Those controls can increase throughput, but none makes prompt injection, dependency risk or architectural mistakes disappear.

The rough week is therefore not an embarrassing footnote to the productivity story. It is the cost curve becoming visible.

July: the project becomes an institution

On 8 July, OpenClaw announced the OpenClaw Foundation as a United States 501(c)(3) non-profit. The announcement introduced the project’s first full-time staff across engineering and operations, said the Foundation would protect the MIT-licensed project’s independence, and listed support from OpenAI, NVIDIA, Microsoft, the University of Michigan, Red Hat and other partners.

Peter remained the principal technical decision-maker. Inside OpenAI, the announcement said, he led Claw Labs on shared product improvements. The Foundation supplied governance, funding and paid maintainership around the open project.

That is the endpoint of the 12-month evolution:

• July 2025: a returned founder learning to supervise a few coding agents;
• October: three to eight Codex sessions, atomic agent commits and continuous refactoring;
• November–December: a burst of reusable tools and a one-hour WhatsApp relay that grew into a personal-agent platform;
• January: viral adoption, thousands of incoming problems and the limits of solo maintenance;
• February: an OpenAI role paired with a promise of independent stewardship;
• May: public acknowledgement that founder-driven release operations had failed users;
• July 2026: a foundation, a full-time team and a project operating at institutional scale.

The individual did not disappear. The system around him changed shape.

So what actually caused the output?

Our reading of the public record points to several causes, not one miracle prompt.

Start with experience. Steinberger brought product instincts, systems knowledge and the scars of managing a large engineering organisation before he opened Codex. The agents multiplied those capabilities; they did not create them.

Then the models crossed his personal threshold for useful delegation. His August workflow still used Claude Code and frequent steering. By October, GPT-5 Codex was the workhorse. By December, he expected many tasks to work on the first attempt. Delegation had become cheap enough to be the default.

Parallelism changed the job again. With three to eight sessions active, inference happened concurrently and waiting became scheduling. Steinberger spent more time deciding, steering and integrating.

He also rebuilt the environment around the agents. CLIs, local tests, docs, screenshots, logs, shared instructions and examples from sibling repositories shortened the path from a prompt to evidence. Oracle handled escalations. CodexBar exposed limits. RepoBar exposed project state. Peekaboo exposed the screen. gogcli exposed services. Each tool made the next one cheaper.

Finally, he tolerated cheap imperfection and refactored continuously. In solo work he committed to main, avoided elaborate branch choreography and asked agents to repair forward. That cut coordination overhead. It also created risks that became harder to tolerate once OpenClaw served a massive community.

The result is not a recipe every team should copy. Committing directly to main behaves differently when one person controls the architecture than when hundreds of contributors depend on stable releases. Giving an agent broad access to a personal computer creates serious security consequences. High commit frequency can hide churn as easily as progress.

But the activity does falsify one comfortable assumption: AI coding agents are only autocomplete with a chat box. In the hands of an experienced operator, with a codebase and toolchain designed around their limitations, they can produce a software-development event stream that looks organisational in scale.

The number and the story

At 64,048 public commit contributions, precision becomes almost absurd. The count can move when a backdated commit reaches a default branch, when GitHub refreshes attribution or when the partial current day advances. GitHub also displayed 198,789 restricted contributions for the account during the period, but exposed no repository or activity breakdown. We excluded them entirely.

That leaves a defensible conclusion.

Over 12 months, Peter Steinberger did not simply write more code. He changed the unit of work. A commit became the residue of a conversation with an agent. A pull request became both an implementation and a statement of intent. A small repository became disposable infrastructure for the next idea. OpenClaw became the place where those tools, methods and ambitions collided with a global contributor base.

The astonishing part is not that one account accumulated 64,048 credited commit contributions.

It is that the year began with a man rediscovering why he wanted to build—and ended with the one-person factory forced to become an institution.

Methodology: OpenClaw Academy queried GitHub on 23 July 2026 at 07:19 UTC. Commit, PR, PR-review and issue headline totals use the GraphQL ContributionsCollection for steipete; the query ended at 2026-07-23T00:00:00Z, but GitHub’s day-bucketed response included the partial 23 July activity visible at retrieval. GitHub’s collection-level fields report the total repositories with commit and opened-PR contributions, while named per-repository breakdowns are capped at 100 for each activity type. GitHub counts commits associated with the account on a repository’s default or gh-pages branch and excludes unmerged fork-only commits; email association and later repository activity can affect attribution. Repository creation counts include original non-forks under @steipete, whether polished, experimental, archived, documentation or asset repositories. The inspected page reached repositories created before the measurement boundary. PR Search totals use author:steipete created:2025-07-23..2026-07-22; the final partial day is excluded because Search qualifiers operate by date. OpenClaw release, contributor and default-branch commit totals were fully paginated through GitHub’s REST API. We did not subtract bot activity from repository-wide OpenClaw totals. Peter’s account totals include agent-generated or automated work committed or opened under his identity. Known mirror totals were not manually added; GitHub’s own public attribution may still include mirrored repositories. Private and restricted contributions were excluded. The reproducible evidence packet is stored with the publication at .editorial/research/peter-steinberger-12-months-2026-07-23.json.