Skip to main content
株式会社オブライト
Software Development2026-07-11

Mosaic (mosaic.inc) Deep Dive — Run Many Claude Code / Codex / OpenCode Sessions in Parallel From One Desktop, Plus SHARED CONTROL for Multiple People Live-Editing a Single Claude Code Session, With Git-Worktree-Isolated Workspaces That Never Collide — a New Option Alongside Conductor / Superset / Herdr in the Late-2026 Parallel-Agent Ops Stack

Mosaic (mosaic.inc) is a desktop tool that runs multiple AI coding agents (Claude Code / Codex / OpenCode) in parallel from a single UI, and on top of that offers "SHARED CONTROL" — a rare feature in 2026 that lets multiple people live-edit a single Claude Code session together. The core of SHARED CONTROL: spin up a session, a teammate joins the same terminal, and you both see each others cursors, prompts, and output in real time — genuinely shared control, not screen-share theater. The experience of Google Docs or VS Code Live Share, applied to a Claude Code session. Use cases (see the standing request in Anthropic issue #60082): pair programming, team debugging, onboarding, live review, and division of labor (Ali refactors the auth middleware while Shubham adds the rate-limit test — same session). Parallel-ops layer: each agent gets a Git-worktree-isolated workspace, and 5-10 agents can develop different branches / features concurrently with zero file collisions or state interference. Position: in the rapidly maturing late-2026 parallel-Claude-Code market, Mosaic competes with Conductor (YC, Mac-native), Superset (three ex-YC CTOs, 10+ concurrent), Herdr (terminal-native, Rust), and Product-Hunt entries like Claudy / Multi-Claude — its distinctive differentiation is real-time SHARED CONTROL. Target users: solo developers, startup CTOs, staff engineers, plus teams that want pair programming and live review with AI agents in the loop. Combined with Command Code taste-1 learning, the grill-me skill, and Crit.md HITL review, it forms a "parallel × personalized × thoroughly reviewed × real-time collaborative" late-2026 AI dev stack.


TL;DR — What Mosaic Is

Mosaic (mosaic.inc) is a desktop tool that runs multiple AI coding agents (Claude Code / Codex / OpenCode, etc.) in parallel from a single UI.

Four takeaways:

1. SHARED CONTROL — multiple people live-edit a single Claude Code session together (a genuine differentiator in 2026)
2. Git-worktree-isolated workspaces — every agent works on a separate branch / feature, zero file collisions
3. 5-10 agents in parallel + GUI desktop experience (contrast with terminal-native Herdr)
4. Claude Code / Codex / OpenCode integrated — cross-agent management

SHARED CONTROL — the Standout Feature: Multiple People, One Session

Mosaic’s biggest differentiator, a feature almost no other tool (Conductor / Superset / Herdr) offers:

> Multiple people, one session.
> Spin up a session, a teammate joins, and you’re both in the same terminal, live. They type, you type, you both see each other’s cursors and prompts as they happen. Real shared control, not screen-share theater.

What it looks like:

one terminal · multiple people
› refactor the auth middleware   Ali
› add the rate limit test         Shubham
✓ both editing, live, same session

The core experience:
- The developer who opened the session shares an entry point; teammates join with mosaic join <session-id> or similar
- Each person’s cursor, prompt, and agent output shows in real time on every screen
- One person can type instructions while the other watches and adds more
- To Claude Code / Codex, it’s one unified conversation — but the tool tracks who said what

Contrast with prior approaches:

PriorMosaic SHARED CONTROL
Zoom / Meet screen shareReal shared control — both people actually type
tmux + SSH (single user)Different users legitimately co-present, permissions separated
Claude Code share linksTwo-way editing, not read-only
session-bridge (file-based)Real-time sync
GitHub PR reviewCollaborate during generation, not after

Background: this is the “Google Docs / VS Code Live Share for Claude Code” that’s been long-requested in Anthropic issue #60082. Anthropic’s native product doesn’t offer it yet — Mosaic ships it early.

Use Cases for SHARED CONTROL

(1) Pair programming: two people in the same Claude Code session — one drives the spec, the other adds test strategy, and Claude integrates both inputs into one code generation.

(2) Team debugging: during a production incident, multiple engineers join one session, each injecting their domain knowledge into Claude to narrow down root cause.

(3) Onboarding: a senior engineer joins the newcomer’s session, watches how they instruct Claude in real time, adds course-correcting nudges — interactive OJT.

(4) Distributed team pairing across time zones: async but same-session — the morning shift picks up the exact Claude conversation the evening shift left off.

(5) Split work inside one session: the tagline example — Ali refactors the auth middleware while Shubham adds the rate-limit test, same session, different subtasks, Claude keeps both contexts and proposes merges.

(6) Live review: skip the async PR loop (via Crit.md or GitHub); the reviewer joins the ongoing Claude Code session and adds fixes as they happen — the review step effectively vanishes.

Considerations When Using SHARED CONTROL

(1) Permission management: multiple people can instruct Claude — you need audit logs of “who said what, when” for enterprise adoption.

(2) Intent collisions: if Ali is telling Claude “adopt Zod” and Shubham says “use Yup instead,” Claude gets confused — pre-agreed role split is a real-world necessity.

(3) LLM token consumption: concurrent prompting accelerates spend against Claude Code Pro / Max limits.

(4) Sync latency: remote-participant cursor sync typically rides WebSocket / WebRTC — network quality dependent.

(5) Security: session join URLs / tokens are the whole ballgame — a leak = session takeover. Be careful sharing them over Slack DMs.

The Problem — the Parallel-Agent Ops Wall

Late 2026 makes "multi-agent parallelism" a practical problem:

- One dev wants 5–10 agents at once — frontend, backend, infra, tests, docs all in parallel
- Same repo, multiple agents collide — the same file gets edited two ways
- Hard to see who's doing what — juggling many terminal windows has a ceiling
- Claude Code's built-in multi-session tops out at 25 parallel, but its UI integration is limited

Mosaic's answer: give each agent a fully isolated Git worktree, and unify state / switching / management in a desktop GUI.

Git-Worktree-Based Isolation

Git worktrees let you check out multiple branches of the same repo into separate directories on disk. Because each branch lives in its own physical directory, concurrent edits don't collide.

Mosaic's implementation:

- Creating a new agent auto-creates a Git worktree + new branch
- The agent operates fully inside that worktree — never touches sibling worktrees
- On completion, review → one-click merge or discard

This is the industry pattern: Conductor / Superset / Anthropic's official Claude Code parallel sessions all use worktrees. Mosaic's take: UI / UX polish and cross-agent-type coverage.

Key Features

1. Desktop-App Unified Management

Every agent in one window:
- Sidebar: all agents listed with state (working / blocked / done / idle)
- Main pane: conversation with the selected agent
- History panel: session history and work logs

2. Full Isolation via Git Worktree

Per-agent workspace:
- File edits / runs / tests happen in separate directories
- No cross-branch commit collisions, no state interference
- Integrate finished work via git merge / cherry-pick

3. Multi-Agent Integration

Cross-CLI-agent support:
- Claude Code — Anthropic-native
- Codex — OpenAI-native (the ChatGPT Work substrate)
- OpenCode — open source
- Other CLI agents

Assign the best-fit agent per task: e.g. Claude Code for new features, Codex for refactors, OpenCode for tests.

4. At-a-Glance Agent State

Sidebar states:
- working — executing
- blocked — waiting on you (approval / answer / review)
- done — finished, awaits your check
- idle — waiting

Instant "who's next?" decisions — same design lineage as Herdr's sidebar.

5. Per-Session Chat and Work Logs

Persistent management:
- Per-agent conversation history stored separately
- Logs of executed commands, edited files, generated code
- "Wait, what did this agent do again?" stays traceable weeks later

Position in the Parallel-Agent-Ops Market

Fast-maturing late-2026 market:

ProductFormNotes
ConductorMac desktopYC-backed, Claude Code + Codex, macOS all-in
SupersetDesktop IDEThree ex-YC CTOs, 10+ parallel, unified dashboard
Mosaic (this column)DesktopMulti-agent, Git-worktree-based
HerdrTerminal (Rust)tmux lineage, SSH, mobile UI
ClaudymacOS wrapperClaude Code-focused, Draft Commits, Marketplace
Multi-ClaudemacOSMultiple Claude accounts side-by-side
ReplicasCloudIsolated cloud VMs
coder/muxDesktopOpen source, isolated parallel dev
HumanLayer CodeLayerCloud + localWorktrees + remote workers, YC-startup adoption

Mosaic's spot: GUI desktop + Claude Code / Codex / OpenCode + Git worktrees — cross-platform-leaning versus Conductor's macOS-native pitch, lightweight-simple versus Superset's full IDE.

Vs Anthropic's Native Parallelism

- Subagents — define specialist subagents with tool-permission control
- Agent View — dispatch sessions, attach when needed
- Agent Teams — a lead + teammates, assigned tasks and review
- Dynamic Workflows — spawn and cross-verify multiple subagents
- Up to 25 concurrent threads

Why external tools like Mosaic remain necessary:
- Native features handle inside-Claude-Code parallelism — not Codex / OpenCode
- GUI dashboard polish is ahead in third-party tools
- There's a niche between CLI-native and full desktop
- Team / project-management integration goes further outside the native product

How People Use It

(1) Solo dev, parallel projects: three worktrees / three agents for front / back / infra.

(2) Startup CTO high-throughput features: 5–10 features handed to 5–10 agents, standup checks each morning.

(3) Large refactors: different agents on different monorepo packages — Codex on legacy, Claude Code on new work.

(4) Experimentation vs mainline: run risky designs in a separate worktree while another agent keeps the main line moving.

(5) OSS maintenance: parallel handling of many issues / PRs, all visible from Mosaic.

How It Composes With Other Tools

The ideal late-2026 AI dev stack:

LayerTool
Parallel-execution substrateMosaic (this column) or Herdr
AgentsClaude Code / Codex / Cursor
PersonalizationCommand Code's taste-1
Design phaseThe grill-me skill (Anthropic Claude Skills)
Diff reviewHunk (terminal) or Crit.md (browser)
Model substrateNous Portal or direct APIs
InfrastructureCloudflare-only stack

Mosaic is the GUI-fan choice at the parallel-execution layer, CLI fans go Herdr, macOS-devoted teams go Conductor.

Caveats and Warnings

(1) LLM API costs balloon: 5–10 agents in parallel run up token spend fast; pair with Claude Sonnet 5 / Opus 4.8 pricing and set budget alerts.

(2) Worktree disk cost: 10 concurrent worktrees can eat significant disk; clean up regularly.

(3) Merge conflicts still happen: parallelism prevents them mid-run, but when it's time to merge to main they resurface — conflict-resolution skill remains required.

(4) Cognitive load: managing 10 agents at once is its own kind of fatigue — a different flavor than grill-me's question fatigue.

(5) Keeping up with agent APIs: Claude Code / Codex / OpenCode change; the tool has to track them — community activity matters long-term.

(6) Production governance: personal use is easy; org-wide multi-agent operations need access control, audit logs, and API-key management designed for it.

Recommended Actions

Solo developers: try Mosaic + Command Code for parallel × personalization; combine with the Claude Code Pro plan to feel 5–10 agents in parallel.

Startup CTOs: use Mosaic to develop features in parallel, Crit.md for review, and grill-me for design — the three-layer setup unlocks high-throughput implementation.

Teams: personal PoC first, then evaluate access control and audit logging needs before scaling. Ask the vendor about an Enterprise SKU.

Bottom Line

Mosaic is the GUI-desktop option in the late-2026 parallel Claude Code / Codex / OpenCode ops market, with Git-worktree-isolated workspaces letting 5–10 agents run concurrently without collisions. Competing with Conductor (macOS-native), Superset (full IDE), Herdr (terminal), and Product-Hunt-launched Claudy / Multi-Claude, its differentiation is GUI desktop + multi-agent-type integration + Git worktrees. Combined with Command Code's taste-1 learning, grill-me, Crit.md HITL review, and the Hunk diff viewer, it becomes the parallel-execution substrate for a "parallel × personalized × thoroughly reviewed" late-2026 AI dev stack. Even as Anthropic's native Subagents / Agent View / Agent Teams / Dynamic Workflows expand, GUI dashboards + cross-agent-type coverage keeps demand for outside tools. Six caveats: rapid LLM-API cost growth, worktree disk footprint, merge conflicts at integration time, cognitive load, keeping pace with agent APIs, and org-level governance.

Related services from us — software development, AI consulting, Hermes Agent setup, and OpenClaw setup. For enterprise adoption of parallel Claude Code / Codex operations, Git-worktree-based workflow design, or choosing between Mosaic / Conductor / Superset, get in touch.

References

Feel free to contact us

Contact Us