Mosaic, a San Francisco startup in Y Combinator's Summer 2026 batch, says about 2,000 people have installed its software, which syncs Claude Code, Codex and Cursor sessions into one shared store. Thirty-seven organizations were actively syncing at launch, according to the company's YC Launch post.
The timing is hardly accidental. Salesforce shipped Slack Code in late August 2026, and YC added "Multiplayer AI" to its Fall 2026 Request for Startups. Team-based agent coding is now a funded category with large-company entrants.
What Mosaic does
The company describes its product as "shared memory for your team's agents." Per the launch post, the idea is that any teammate or agent can pick up "where someone else left off," with session state carried across people, tools and time.
Mosaic's public docs say the install is macOS only. It runs through Homebrew or a shell installer, registers a LaunchAgent and keeps a read-only local replica of the team's data at ~/.mosaic. It also ships "skills" into supported agents.
The command-line tool has three notable commands: mosaic fork continues a shared session in its native agent, mosaic handoff assigns work to a teammate, and mosaic scope controls what gets backed up.
The adoption figures come from Mosaic itself. Wavect's Sept. 7 buyer review repeated them as company-reported and unaudited. Since the launch post is now a couple of months old, anyone evaluating the product would be wise to ask for current numbers.
A two-person team
Mosaic has two people. Its YC page lists Shubham Patil, a former founding engineer at Beem, and Dorsa Rohani, whose background includes NVIDIA, Tenstorrent and Deep Genomics. Rohani was also named to the MLH Top 50.
A June 2026 company memo lays out the thesis. "Two agents with identical context are the same agent," it says, adding: "The real unit isn't the agent, it's one shared context store every agent reads from and writes to."
Why multiplayer is arriving now
Aaron Epstein, a YC partner, put the case in the Fall 2026 RFS: "Anyone on a team should be able to drop into the same live agent session to watch it work, redirect it, and hand it off… Anywhere a team already crowds around one problem, there should be multiplayer agents they all share."
Salesforce, for its part, said Slack Code, launched Aug. 24-25, aims to "make AI software development multiplayer." The product offers shared code channels where cross-functional teams work alongside agents such as Claude Code, Devin, GitHub Copilot, ChatGPT and Vercel agents, across all Slack plans. Salesforce and Anthropic said on Aug. 26 that Anthropic is a founding partner. On Oct. 1, Rob Seaman, a Salesforce executive vice president and the general manager of Slack, put it bluntly: "Agentic coding is now multiplayer."
Anthropic's June 24 guidance on human-agent teams makes the memory argument directly. "Agents build their understanding entirely from the text a team makes searchable," it says. YC, meanwhile, open-sourced "qm" on July 29, a multi-user agent harness for Slack and the web with per-user memory and sandboxes.
Usage data points to real demand. The Stack Overflow Developer Survey 2026, published Oct. 6, found that 73% of respondents use coding AI tools daily. Gartner forecast on July 20 that spending on AI platforms and models will grow 63% in 2026.
Speed, though, has not solved everything. In an Atlassian survey reported in October, 87% of developers said AI helps them write code faster, yet 72% reported the same or greater friction across the full software development lifecycle.
A crowded field
Mosaic is far from alone in shared memory. Gryz, Universal Memory, Dory, AO2 Memory, Stele, Hydrate, PLUR, Agent Brain, Myrmo and Statefulai all market similar layers, according to their own sites.
Live-collaboration tools come at the problem from another angle. Superconductor, Conductor, Amp, HumanLayer and Agent FM (also YC S26) offer shared workspaces or control planes, and Anthropic's Claude Code has its own Agent Teams feature with shared task lists and messaging.
Poly, a rival, published a comparison on Oct. 3 that characterizes Mosaic as cross-session, read-only shared history, set against Poly's single live agent room with approvals and cost receipts. Poly also said Mosaic's pricing was not published. The comparison is vendor-authored, so readers should weigh it accordingly.
Governance is the open question
A central store of session transcripts is a tempting target for attackers. Wavect's review cautioned that synced transcripts do not amount to trustworthy memory, and it told buyers to assess provenance, scope, handling of stale context, redaction, deletion, export and permission inheritance.
As of Oct. 27, Wavect said, Mosaic's public docs lacked detail on tenant isolation, encryption, SSO, audit logs, stale-context invalidation, redaction, data-region controls and pricing.
Outside pressure is building. Black Duck and UserEvidence reported in March that only 30% of organizations have full governance in place, even as 78% of heavy AI coding-assistant users reported major productivity gains. The EU AI Act's enforcement powers became active Aug. 2, 2026.
Cost is another worry. ITPro and TechRadar coverage in October described enterprises caught off guard by surprise AI bills from consumption pricing and agent workloads.
What to watch
Mosaic has to show it can answer the security questions buyers are now asking, while competing against Slack's distribution and a dozen similar memory tools. As of Oct. 27, third-party reviews noted that Mosaic had not published pricing and that its public docs lacked detail on security specifics such as tenant isolation, encryption, SSO, audit logs and data-region controls.
One wildcard is the Linux Foundation's Agentic AI Foundation, announced Dec. 9, 2025, with Anthropic's Model Context Protocol among its founding projects. It may set the interoperability baseline these tools build on. For Mosaic, the nearer test is simpler: whether 37 organizations turns into a number the company can keep updating.
