Screenpipe, a Y Combinator Summer 2026 startup, has released a beta of Workflows, a feature that converts recorded on-screen work into reviewable workflow drafts and editable agent skills, according to the company's September 16 announcement. The product targets a problem enterprise buyers know well: before a task can be automated, someone has to document it.
Screenpipe's pitch is that watching the work can replace the documenting. Much of the evidence for that pitch, though, comes from the company itself.
What Workflows does
According to the company's Workflows announcement (Screenpipe #32), the beta mainly turns observed work into context, workflow drafts, standard operating procedures and editable agent skills. It is not automatic foundation-model training on every employee's activity.
The announcement's example follows a support task from Intercom to Linear to Slack. The data in that example is fictional, and the company has not described it as a customer deployment.
A September 23 update (Screenpipe #33) added chat and workflow guides. Those materials call Workflows a beta or private beta, while the company's current enterprise marketing reads as broader. No general availability date has been given.
How the enterprise pitch works
The enterprise page describes AI-led interviews meant to capture the reasoning, exceptions and checks that screen recordings cannot show. It also describes testing an agent as it repeats a workflow in a separate workspace, with human review before anything reaches production.
Companies can optionally train models on approved workflows, with the scope agreed in advance. All of this is the company's own description; the research for this article turned up no independent benchmark of the end-to-end process.
In a Launch HN post dated July 23, a founder said an authenticated local API, MCP and skills expose recorded history to agents.
Privacy claims need qualifiers
Screenpipe's privacy policy, updated October 10, says recordings, transcripts and pipe data stay on the device in local-only mode. Data can leave the device when users turn on cloud AI, sync, exports, connectors, team workflows or support features. The policy names Negentropy Labs, Inc., doing business as Screenpipe, as the operator.
The GitHub README describes optional encryption at rest, optional encrypted device sync, support for local models and per-pipe data-access controls. It adds that transcription, integrations, sync and telemetry must be configured separately to control network behavior. A flat claim that nothing ever leaves the computer would therefore be inaccurate.
The enterprise FAQ says employees using personal capture can review outputs before sharing them. It also warns that organization-managed storage and policies can permit broader access.
Reviewers have raised concerns. Anchor Terminal reported default analytics, error-reporting and support-log pathways, and criticized earlier absolute privacy language; its findings rest on its own review dates. In the July Hacker News thread, one user alleged that an automation sent local API keys to an endpoint. Founder Louis Beaumont offered configuration help and said filtering was on by default. Nobody verified a reproduction or a resolution, so the matter stands as an unresolved allegation rather than a confirmed breach.
Screenpipe's trust center, checked November 8, labels the company compliant with SOC 2 Type 2, ISO 27001, GDPR, HIPAA and CCPA, and lists EU AI Act readiness as in progress. The public page did not offer audit reports for inspection.
Traction and licensing
The enterprise page, observed November 8, claims more than 500 customers and usage within the Fortune 300. It does not name the Fortune 300 organization or define what counts as a customer. The testimonials on the page were published by the vendor.
A Dealroom note attributes to Beaumont a claim, made at YC Demo Day on September 10, that revenue grew tenfold during the program. The note gives no absolute revenue figure or revenue basis, and the underlying X post could not be independently retrieved.
Licensing is murkier than older coverage suggests. Screenpipe has published a license update, and the research indicates that historical articles calling the product simply "open source" need qualification. Its current onboarding page says shared context across teammates is not included in the Business tier, and that commercial source builds require Enterprise and a commercial license.
A Master Services Agreement dated November 4 covers desktop software, hosted functionality, work-activity analysis, workflow documentation and task-execution assistance. It is a review template, not evidence of a signed customer deal.
Funding and founders
YC confirmed Screenpipe's Summer 2026 participation. Caplight lists an accelerator round dated May 14, 2026, with $500,000 in total funding, but that is a database figure and not a company-confirmed total.
Headcount is similarly fuzzy. YC's profile lists Beaumont alone, with a separate record showing two people, while LinkedIn shows a range of 2 to 10. The company has not disclosed an exact number.
Beaumont previously worked on Mediar with Matthew Diakonov. Investor Choongsoon Bae referred to "Mediar (Screenpipe)" in a post about a July 2025 financing. The research does not establish how that $2.8 million round, or a $40 million valuation, maps to the current Negentropy Labs entity. Neither should be read as a new 2026 round or a current valuation.
Who else is in the space
Mimica positions itself as "AI agents that watch & learn," covering process capture, mapping and automation. Scribe sells Capture and Optimize products, and a September 6 post from its CEO also names Agents and MCP. UiPath Task Mining records task traces and variations to identify bottlenecks before exporting them into automation or documentation workflows.
Microsoft Recall sits nearby but is a personal-history product, not an enterprise agent-building platform. Microsoft says Recall snapshots are stored and processed locally, encrypted, and accessed with Windows Hello, and that the feature does not record audio.
What buyers should ask for
Public customer evidence is mostly company-published testimonials. Per the research, security teams should request audit reports, deployment-specific data-flow diagrams, redaction evaluations, retention and deletion details, and documentation of administrator access. Operations leaders should ask for named deployments, completion and error rates, and how often humans have to step in.
Screenpipe's newest newsletter visible on November 8 (Screenpipe #36) described starring a stretch of work and reusing its context with an AI assistant or automation. Whether Workflows reaches general availability is the next thing to watch.
