Y Combinator, the startup accelerator behind Airbnb, Stripe, and Reddit, just open-sourced the AI system that runs its entire company. QM, a multiplayer agent harness, handles accounting, legal, events, and engineering at YC. It even helped build itself.
TL;DR: Y Combinator open-sourced QM under the MIT license — a multi-agent harness used internally across accounting, legal, events, and engineering teams. The tool provides scoped workspaces, shared memory, Slack integration, and a web UI. Eve, YC’s top power user, describes the lean team outputting “like an army” with QM (Eve Bouff, X, 2026).
What Is QM and Why Did Y Combinator Build It?
QM is a multiplayer agent harness designed for company-wide AI collaboration, released under the MIT license. According to Y Combinator’s announcement, the tool is “meant to be easy to customize, like Hermes or OpenClaw, but useful for a whole company” (Y Combinator, X, 2026). Unlike single-purpose agent tools, QM targets organizational workflows from day one.
The core problem QM solves is isolation. Most agent frameworks treat each session as a standalone interaction. QM gives every employee their own scoped memory, files, permissions, and sandbox while enabling collaboration through shared channels and projects. This dual model — personal workspace plus team coordination — sets QM apart from existing tools.
Y Combinator describes itself as “an intentionally, ridiculously lean team” that outputs like a much larger organization (Eve Bouff, X, 2026). QM plays a direct role in that leverage. The harness runs agent workflows across multiple departments simultaneously, with shared memory and browser support baked in. It is cloud-first, with Slack and web UI as native interfaces rather than afterthoughts.
The project is not experimental. YC uses QM in production for real business operations.
How Does QM’s Multiplayer Agent Model Actually Work?
QM’s architecture centers on scoped workspaces that combine personal isolation with collaborative channels. Each person gets isolated memory, file access, and a sandbox environment. Meanwhile, shared Slack channels, group messages, and projects allow multiple agents and humans to interact in the same context (Vaibhav Sisinty, X, 2026).
The multiplayer model means agents operate as team members, not solitary tools. Work happens in channels where humans and agents communicate together. This mirrors how real teams already function in Slack. The difference is that AI agents participate as active contributors with persistent context.
Key architectural features include:
- Company scopes — organize agents, permissions, and data by organizational boundaries
- Crons — scheduled agent tasks that run automatically on defined intervals
- Skills — modular capabilities agents can invoke, similar to function calling or tools
- Shared memory — persistent context across sessions, channels, and team members
- Browser support — agents can navigate and interact with web content
- Sandboxed execution — isolated file systems and permissions per user workspace
- Slack-native integration — agents live in channels alongside human colleagues
- Web UI — a dedicated interface for direct agent interaction outside Slack
| Feature | QM | Hermes | OpenClaw |
|---|---|---|---|
| License | MIT | Open source | Open source |
| Target | Whole company | Individual/team | Individual/team |
| Slack native | Yes | Limited | Limited |
| Web UI | Yes | Varies | Varies |
| Scoped memory | Per user | Per session | Per session |
The combination of crons and skills gives QM a workflow-engine quality. Agents do not just respond to prompts — they execute scheduled tasks and chain modular capabilities.
Which Departments at YC Use QM in Production?
Y Combinator runs QM across four departments: accounting, legal, events, and engineering (Y Combinator, X, 2026). Notably, the engineering team uses QM to build QM itself, making the harness self-hosting in its development cycle.
This is not a pilot program. YC staff and partners use QM daily for real work. Eve, described as the “#1 power user” of QM internally, states that watching YC partners and staff use the tool to “1000x themselves” has been “mesmerizing” (Eve Bouff, X, 2026). The language is strong, but it reflects actual production usage across business-critical functions.
The departmental spread matters. Accounting involves structured data, compliance, and reporting. Legal requires document analysis, contract review, and risk assessment. Events demand logistics coordination, scheduling, and communication. Engineering needs code generation, review, and system architecture. QM serves all four without department-specific customization — the harness adapts through its skill system and scoped workspaces.
Y Combinator’s lean structure makes this relevant. The accelerator operates with a small team relative to its portfolio size and program scope. QM fills the gap between headcount and workload.
How Does QM Compare to Hermes and OpenClaw?
Y Combinator explicitly positions QM alongside Hermes and OpenClaw in its announcement. The key differentiator is scope. Hermes and OpenClaw target individual developers or small teams. QM targets entire organizations (Y Combinator, X, 2026).
The comparison is deliberate. Hermes and OpenClaw established the pattern of customizable agent harnesses with pluggable backends. QM adopts that flexibility but adds organizational layers — company scopes, per-user workspaces, shared memory across teams, and Slack-native collaboration. Where Hermes might serve a single developer’s coding workflow, QM serves accounting, legal, events, and engineering simultaneously.
Elvis, commenting on the release, noted that “lots of agent work is now happening with an agent team in channels/groups” and confirmed QM works with Pi, OpenCode, Claude Code, and Codex as backends (Elvis, X, 2026). This backend flexibility means QM does not lock users into a single model provider. Teams can route different tasks to different models.
What Architecture and Backends Does QM Support?
QM runs on a single core with four backend options, according to AI/TLDR’s coverage of the release (AI/TLDR, 2026). The supported backends include Pi, OpenCode, Claude Code, and Codex. This multi-backend design lets organizations choose models based on task requirements, cost, or existing contracts.
The architecture is cloud-first by design. YC built QM for environments where teams collaborate remotely and agents need persistent access to shared resources. The cloud-first approach means QM does not assume local-only execution or single-machine deployments. Instead, it expects distributed access patterns from the start.
Native integrations cover two primary interfaces: Slack and a web UI. Slack integration means agents participate in existing communication channels. The web UI provides a direct interaction surface for tasks that do not fit naturally into chat. Both interfaces connect to the same underlying workspace, memory, and skill system.
The shared memory model persists across sessions and interfaces. An agent that learns something in a Slack channel can apply that knowledge in a web UI session. This continuity is what makes QM function as an organizational tool rather than a conversation partner. Browser support extends agent capabilities beyond text — agents can read documentation, interact with web apps, and gather information from live sources.
How Do Scopes, Permissions, and Sandboxes Function?
QM assigns every employee a scoped environment with isolated memory, files, permissions, and a dedicated sandbox, according to details shared by YC power users (X/@VaibhavSisinty, 2026). Each person operates within boundaries that prevent accidental interference with another employee’s agent workspace or data.
The architecture treats isolation as a first-class concern. Agents cannot cross scope boundaries without explicit project-level collaboration invitations. This design prevents a billing agent from accessing legal documents or an engineering agent from touching accounting records.
Sandboxed execution means agents run code in contained environments. File system access stays within the assigned scope. Network calls follow permission rules set at the company level. The system enforces these boundaries at the harness layer, not at the agent level.
Shared memory operates differently. When employees join a project or a Slack channel, their agents gain access to that shared context. Leave the channel, and the access disappears. This scoped sharing model mirrors how human teams actually collaborate — you see what you’re invited to, nothing more.
Permission granularity extends to skill execution. A company administrator defines which skills each scope can invoke. For example, an accounting team might have browser automation enabled while a legal team gets document analysis tools. These configurations live in declarative files that version-control cleanly.
What Can Companies Customize in QM?
Companies can customize skills, cron schedules, backend providers, and project-level agent configurations, as documented in the QM repository and confirmed by explainx.ai (2026). The harness ships with a modular core designed for extension rather than replacement.
Skills represent the primary customization surface. A skill is a discrete capability — browser automation, document parsing, code execution — that an agent can invoke. Organizations write custom skills in standard code and register them with the harness. The four-backend architecture (supporting Pi, OpenCode, Claude Code, and Codex) means companies choose their preferred model provider.
Cron schedules enable automated agent workflows. A company might configure a weekly accounting reconciliation agent, a daily legal document review, or hourly engineering test runs. These scheduled agents operate within their assigned scopes and follow the same permission rules as interactive sessions.
Project-level configuration lets teams define agent behavior per project. One project might use Claude Code for code generation tasks. Another might rely on Codex for different work. The harness abstracts these choices into configuration files.
Customization extends to the UI layer. The web interface supports theming and layout adjustments. Slack integration patterns — which channels agents monitor, how they respond, what triggers invocation — are all configurable.
How Does QM Integrate With Slack and Web Interfaces?
QM ships with native Slack and web UI support from day one, according to Y Combinator’s official announcement (X/@ycombinator, 2026). No additional connectors or middleware are required to get agents running in either environment.
Slack integration turns channels into agent collaboration spaces. Employees mention an agent in a channel, and it responds within that channel’s shared context. Group messages work the same way. Multiple employees can interact with a single agent simultaneously, and the agent maintains coherent state across all participants.
The web interface provides a richer interaction surface. It supports file uploads, long-form conversations, project management views, and detailed agent output inspection. For engineering work — including building QM itself, as YC notes — the web UI handles code review, diff visualization, and multi-file editing scenarios.
Both interfaces connect to the same underlying harness. An employee might start a conversation in Slack during a quick sync, then switch to the web UI for deeper work. State persists across both surfaces because the harness, not the frontend, manages memory and context.
Channel-based agent work reflects a broader industry trend. As Elvis (X/@omarsar0, 2026) observed, agent teams increasingly operate in channels and groups. QM’s native Slack support positions it squarely in this workflow pattern.
What Does QM’s Open-Source License Mean for Adopters?
QM is released under the MIT license, one of the most permissive open-source licenses available, as confirmed by Y Combinator’s announcement and multiple sources (X/@ycombinator, 2026; explainx.ai, 2026). This means companies can use, modify, distribute, and even sell products built on QM without royalty obligations.
The MIT license imposes only two requirements: retain the copyright notice and include the license text in copies. There are no copyleft restrictions. Companies can build proprietary systems on top of QM without open-sourcing their own modifications.
For adopters, this removes a significant adoption barrier. Many enterprise agent tools carry restrictive licenses or usage-based pricing. QM’s MIT license allows internal deployment without legal review cycles. Teams can fork it, extend it, and ship commercial products without negotiation.
The open-source release also means community contributions. YC has indicated that QM is meant to be easy to customize, like Hermes or OpenClaw. The repository accepts pull requests, and the modular architecture makes contributions straightforward.
| License Feature | QM (MIT) | Typical SaaS Agent Tools |
|---|---|---|
| Commercial use | Permitted | Subscription-based |
| Source modification | Permitted | Closed source |
| Distribution | Permitted | Vendor-controlled |
| Private deployment | Full control | Limited or none |
| Cost | Free | $20–$200 per seat/month |
Who Should Deploy QM and Who Should Wait?
Organizations with existing DevOps capabilities and a clear use case across multiple departments should deploy QM now. Teams without infrastructure experience or those needing a managed solution should wait for ecosystem maturation.
Y Combinator itself uses QM across accounting, legal, events, and engineering — demonstrating that the harness handles diverse, real-world workloads (X/@ycombinator, 2026). Eve, YC’s self-described top power user, noted that the lean YC team outputs like a much larger organization partly because of QM (X/@eve_bouff, 2026).
Companies that should deploy now typically share these characteristics:
- Engineering bandwidth to deploy and maintain a self-hosted agent harness
- Multi-department needs where agents assist accounting, legal, operations, and development teams
- Slack-centric workflows where channel-based collaboration is already the norm
- Familiarity with agent tools like Claude Code, Codex, or OpenCode
- Data sovereignty requirements that mandate on-premise or controlled-cloud deployment
- Desire for customization beyond what commercial agent platforms offer
- Budget constraints that make per-seat SaaS pricing impractical at scale
- Open-source preference for auditability and community-driven improvement
Companies that should wait include those without dedicated technical staff, organizations standardized on Microsoft Teams rather than Slack, and teams that need vendor support contracts. QM is powerful but requires operational ownership.
Frequently Asked Questions
Is QM free to use for commercial projects?
Yes. QM is released under the MIT license, which permits unrestricted commercial use, modification, and distribution (X/@ycombinator, 2026). Companies can build and sell proprietary products on top of QM without any royalty or licensing fees.
Can QM work with coding agents like Claude Code and Codex?
Yes. QM supports four agent backends: Pi, OpenCode, Claude Code, and Codex, as confirmed by both the YC announcement and community analysis (X/@omarsar0, 2026; AI/TLDR, 2026). Projects can configure different backends per use case.
Does QM require a cloud environment or can it run locally?
QM is designed as cloud-first, according to Y Combinator’s official announcement (X/@ycombinator, 2026). While the MIT-licensed code can theoretically be self-hosted on local infrastructure, the architecture assumes cloud deployment with Slack and web UI access.
How does QM handle shared memory between employees?
Each employee receives scoped memory, and shared context is created through project participation and Slack channel membership (X/@VaibhavSisinty, 2026). When an employee leaves a channel or project, their agent’s access to that shared memory is revoked automatically.
Summary
QM represents a proven, production-tested approach to company-wide agent orchestration. Here are the key takeaways:
- MIT-licensed and fully open-source — no commercial restrictions, no per-seat costs, no vendor lock-in
- Multiplayer by design — every employee gets isolated scopes with optional shared collaboration in projects and Slack channels
- Four backend support — works with Pi, OpenCode, Claude Code, and Codex, giving teams model flexibility
- Battle-tested at YC — runs accounting, legal, events, and engineering workloads at one of the most influential startup accelerators in the world
- Cloud-first with native Slack and web UI — designed for real organizational workflows, not experimental prototypes
For teams evaluating agent harnesses, the QM repository and documentation are available now. Read the code, test the deployment, and decide whether your organization fits the profile of a successful QM adopter.