Who is Peter Steinberger? PSPDFKit, OpenClaw, OpenAI, and the new agent stack

Peter Steinberger is a useful founder case study because his career maps two eras of developer tools. PSPDFKit was a deep infrastructure business: a PDF rendering, annotation, signing, and document-processing SDK sold into serious mobile, web, and enterprise software teams. OpenClaw is a different kind of surface: an open-source personal agent gateway that connects models, channels, devices, tools, and coding workflows.

The through-line is operator leverage. PSPDFKit turned a hard technical niche into a durable software company. OpenClaw turns the messy edges around AI agents into an operating layer: chat apps, local gateways, skills, sessions, mobile nodes, code agents, voice, and security defaults.

As of June 8, 2026, Steinberger is at OpenAI working on agents, and OpenClaw remains open-source and independent rather than simply being folded into OpenAI. That distinction matters for founders choosing whether to bet on OpenClaw as infrastructure, study it as a pattern, or stay inside managed tools such as Codex and Claude Code.

Before OpenClaw, Steinberger was the founder most associated with PSPDFKit. The product was not a consumer app. It was developer infrastructure for embedding PDF and document workflows into other products: viewing, annotating, signing, processing, and collaborating on documents across mobile and web applications.

That market sounds narrow until you look at the customer base. PSPDFKit publicly described itself as powering apps used by nearly a billion end users, with customers including Dropbox, DocuSign, SAP, IBM, Volkswagen, Fabasoft, Wolters Kluwer Deutschland, and the European Patent Office.

For founders, the lesson is that infrastructure companies can compound quietly. PSPDFKit did not need to be a flashy social product. It solved a painful, technically demanding workflow for software teams that did not want to build document rendering and manipulation themselves. That is the same founder logic behind many strong stack decisions: choose the boring problem that customers pay to avoid.

The cleanest public wording is "strategic investment" rather than a simple full acquisition. On October 1, 2021, PSPDFKit announced its first strategic growth investment led by Insight Partners, totalling more than EUR 100 million. That money was intended to accelerate product growth and fund acquisitions.

Deal coverage from JUVE added the founder-exit texture: Insight acquired shares in PSPDFKit, the sellers included co-founders Jonathan Rhyne, Peter Steinberger, and Martin Schuerrer, the parties did not disclose the purchase price, and Steinberger and Schuerrer were stepping back operationally while retaining an ongoing stake.

That is why it is fair to describe the PSPDFKit outcome as a successful exit path, but not as "Peter sold 100 percent of the company and disappeared." It was a private-equity growth investment with secondary liquidity, continued company expansion, and later rebranding into Nutrient. For SEO shorthand, "PSPDFKit exit" is understandable; for accuracy, "strategic growth investment with founder liquidity" is closer to the public record.

In October 2024, PSPDFKit rebranded as Nutrient. The rebrand followed the 2021 Insight Partners investment and a roll-up period in which the company acquired ORPALIS, Aquaforest, Muhimbi, and Integrify to build a wider document platform.

Nutrient said the business had tripled revenue since the 2021 investment. That does not tell us Steinberger’s personal exit economics, but it does reinforce the founder trajectory: he helped build a durable developer-infrastructure company that became a larger document automation platform after institutional capital entered.

The product link still matters. If you want to understand the first act, study PSPDFKit/Nutrient as a document SDK and workflow automation platform, not just as a line on a founder biography. The second act with OpenClaw makes more sense when you see the first act as infrastructure, not influencer theatre.

OpenClaw is Steinberger’s second major public success. Its GitHub repository currently positions it as a personal AI assistant you run on your own devices, with a gateway that connects to the channels you already use. The docs describe it as a self-hosted gateway for AI agents across Discord, Google Chat, iMessage, Matrix, Microsoft Teams, Signal, Slack, Telegram, WhatsApp, Zalo, and more.

The important product idea is the gateway. A model by itself answers prompts. A gateway routes messages, manages sessions, connects channels, exposes tools, and gives an agent a place to live outside one chat tab. That is why OpenClaw sits next to Codex and Claude Code in the agent conversation even though it is not another coding model.

OpenClaw also has a developer-operator feel: Node-based install, a daemon mode, web control UI, skills, multi-agent routing, mobile nodes, voice, canvas, and explicit warnings around DM access, sandboxing, and remote exposure. It is not just a demo. It is an attempt to make personal agents operational.

Steinberger’s own February 14, 2026 post is the best primary source. He wrote that he was joining OpenAI to work on bringing agents to everyone, and that OpenClaw would move to a foundation and stay open and independent.

He also framed the decision as a tradeoff. He could imagine OpenClaw becoming a large company, but said he had already done the company-building game with PSPDFKit and wanted to work on broader agent adoption with access to frontier research and models.

So the most careful wording is: OpenAI hired Steinberger, sponsors or supports the project according to Steinberger’s post and OpenClaw’s public repository sponsors section, and OpenClaw is being structured as an independent foundation. Calling it a straight acquisition is too loose unless a later corporate transaction is publicly documented. The better phrase is "OpenAI hired the OpenClaw creator" or, more cautiously, "an OpenAI acqui-hire-style move around the founder, not a confirmed OpenClaw product acquisition."

OpenClaw’s strengths are clear. It is open source, self-hosted, multi-channel, agent-native, and designed around a persistent gateway rather than a single terminal session. Its current repository shows very high community momentum, with hundreds of thousands of stars, tens of thousands of forks, active issues and pull requests, and recent releases.

The product direction is compelling because it answers a real gap. Codex and Claude Code are excellent when the work is inside a repo. OpenClaw becomes interesting when work starts in a message, needs context from a device or channel, branches into a code agent, and then reports back through a human-facing workflow.

The weak spots are the same things that make it powerful: broad permissions, channel sprawl, operational complexity, and a young ecosystem. A self-hosted gateway that touches chat apps, files, shell commands, browsers, mobile nodes, and models has a much larger failure surface than a scoped code agent working inside one repository.

These tools are not interchangeable. The useful comparison is by layer. Codex and Claude Code are primary coding agents. Conductor.build is an orchestration workspace for running coding agents across isolated branches and review flows. OpenClaw and Hermes Agent are broader agent platforms or gateways that can connect models, memory, channels, tools, and background work.

Codex is the clean OpenAI default for codebase work: feature changes, refactors, migrations, reviews, tests, pull requests, and cloud or app-based agent runs. Claude Code remains the strong Anthropic default for developers who like terminal-first agent work and Claude’s coding style.

Conductor.build is not trying to replace Codex or Claude Code. Its docs say it simplifies running Claude Code, Codex, and other agents in parallel, with isolated workspaces, branches, setup, review, checks, PRs, and handoff. That makes it a workflow manager for teams or solo founders running several agents at once.

A realistic founder setup is layered. Use Codex or Claude Code as the worker inside the codebase. Use Conductor.build when you want several workers isolated in branches with a consistent setup, diff review, checks, and PR flow. Use OpenClaw or Hermes Agent when the work begins outside the repo: messages, scheduled tasks, personal assistant workflows, inbox triage, mobile capture, or cross-channel follow-up.

For example, OpenClaw could receive a Telegram message about a bug, gather context, route a scoped coding task to Codex, and report the status back. Conductor could run Codex and Claude Code in separate workspaces against two proposed fixes. Trackk would record the project step, the tool choice, the deployment status, the cost signal, and whether the fix actually moved launch readiness forward.

That is the category shift: the winning system is no longer one magical agent. It is a stack of model, harness, workspace, permissions, review, deployment, and project memory. Founders who treat it as a stack will make better choices than founders who chase whichever demo is loudest this week.

For a solo founder, the safest first stack is boring: Codex or Claude Code for scoped repository work, GitHub for version control, Vercel for deployment if you are in the Next.js ecosystem, and Trackk for launch readiness. Add Conductor.build only when parallel agent branches become a real workflow need.

Add OpenClaw or Hermes Agent after you can describe the trust boundary. If the assistant only reads public issues and drafts summaries, risk is low. If it can message customers, run shell commands, touch production credentials, or control cloud resources, the setup needs allowlists, audit logs, sandboxing, cost limits, and rollback.

That staged approach keeps agentic coding practical. Start with a coding agent, then add a workspace layer, then add a personal-agent gateway. The sequence matters because each layer increases both leverage and blast radius.

Security is the first risk. Agent platforms connect to untrusted messages, local files, shell commands, browsers, GitHub, cloud services, and sometimes payment or customer systems. Prompt injection is no longer theoretical when an inbound message can influence an agent with tools.

Reliability is the second risk. Agents can be impressive and still fail at boring boundaries: stale docs, hidden state, half-applied migrations, bad retries, broken tests, hallucinated APIs, race conditions, and changes that pass locally but fail in production. Parallel agents can also create conflicting fixes that look reasonable in isolation.

Cost and governance are the third risk. Always-on agents can burn tokens, spawn work, preserve too much memory, leak context into the wrong provider, or create audit gaps. Open-source does not remove that burden. It moves more responsibility to the operator.

The first lesson is that unglamorous infrastructure can be a huge business. PSPDFKit solved a hard document problem for developers and enterprise teams, then became a platform large enough to attract a major Insight Partners investment.

The second lesson is that distribution can flip quickly when a founder is operating at the edge of a new platform shift. OpenClaw did not need a traditional SaaS launch playbook to become important; it rode GitHub, social proof, open-source adoption, and a clear product metaphor.

The third lesson is to separate the person from the pattern. Steinberger’s own setup may be unusually intense. Copying the whole operating style blindly is a mistake. The useful pattern is to build tools that remove repeated friction, publish in public, let users remix, and keep the project close to real workflows. That is also the durable SEO lesson: useful founder content should explain the workflow, the tradeoffs, and the source trail, not only repeat the headline.

If you add these tools to a project, track them as an operating system, not as random subscriptions. Create steps for model access, agent harness, workspace isolation, GitHub permissions, environment variable handling, local setup scripts, test commands, browser QA, deployment review, rollback, logs, and spend limits.

For Codex and Claude Code, the core readiness checks are repo instructions, tests, branch discipline, approval boundaries, and diff review. For Conductor.build, add workspace setup, copied config, port strategy, merge rules, and PR ownership. For OpenClaw or Hermes Agent, add channel allowlists, sandboxing, gateway exposure, provider routing, memory policy, and incident recovery.

The founder advantage is not merely "AI writes code." The advantage is a visible operating loop where agents increase throughput while Trackk keeps the quality gates, costs, and launch readiness honest.