Paseo: Remote Agent Orchestration for the Multi-Device Developer
Summary
Architecture & Design
Unified Control Plane
Paseo operates as a transparent proxy between your devices and various AI coding agents. Rather than replacing existing tools, it wraps them in a controllable API with real-time state synchronization, exposing a consistent interface whether you're on a phone, laptop, or headless server.
| Component | Function | Configuration |
|---|---|---|
Paseo Server | Central task queue & session state management | Self-hosted (Docker) or managed cloud; WebSocket/SSE for real-time sync |
Agent Adapters | Provider-specific bridges (Claude, Gemini, Codex, OpenCode) | Env vars for API keys; custom system prompts per adapter |
Mobile Client | iOS/Android app for remote monitoring | Biometric auth, push notifications for agent completion/failure |
CLI Bridge | Local agent proxy with tunneling capabilities | paseo init --agent=claude-code --remote |
Developer Workflow Integration
The workflow follows a session-based model. You initiate an agent locally via paseo run, which transparently tunnels I/O through the orchestration server. This enables the "walk away" pattern—approve destructive git operations from your phone via haptic-confirmed notifications, or restart failed builds without opening your laptop. The paseo attach <session-id> command allows seamless handoff between devices, maintaining full terminal history and context.
Key Innovations
The Mobile Gap Solution
Current AI coding agents are desktop-hostage. Paseo breaks this by treating agents as remote processes you can monitor from anywhere—like SSH for the AI age, but with structured semantics and safety guardrails.
Key Differentiators
- Cross-Device State Persistence: Unlike Claude Code or Aider which terminate when you close your laptop, Paseo maintains agent state server-side, enabling seamless handoff between desktop and mobile without losing context.
- Multi-Agent Orchestration: Run Claude for architecture planning while Codex handles implementation, managed from a single mobile dashboard. Most competing tools lock you into single-provider workflows.
- Async Approval Workflows: Configure agents to pause on destructive operations (git push to main, database migrations) and notify your phone—critical for autonomous agents that might otherwise run unsupervised.
DX Improvements
The mobile interface optimizes for high-signal, low-interaction design: swipe gestures to approve diffs, voice-to-text for quick commands, and smart grouping of agent outputs by file changes. The CLI supports --watch mode for specific files, triggering mobile notifications only when relevant code is modified.
Performance Characteristics
Latency & Resource Footprint
As a thin orchestration layer written in TypeScript, Paseo adds minimal overhead to native agent performance. The server component maintains a ~150MB memory footprint under normal load, with WebSocket connections sustaining sub-100ms latency for UI updates over 4G networks.
| Capability | Paseo | Claude Code | Aider | Continue.dev |
|---|---|---|---|---|
| Cross-Device Access | Native (Mobile/Desktop/CLI) | Terminal only | Terminal only | IDE only |
| Multi-Agent Orchestration | Yes (Simultaneous) | Anthropic only | Single session | Multi-model |
| Remote Monitoring | Core feature | No | No | No |
| Startup Latency | ~800ms (CLI handshake) | ~2s | ~1s | Instant (IDE) |
| Concurrent Sessions | 10+ (tested) | 1 per terminal | Limited by terminal | Tab-based |
Scalability Considerations
The architecture supports multiple concurrent agent sessions through connection pooling, though current limitations emerge at >10 simultaneous agents due to WebSocket overhead. For individual developers and small teams (3-5 agents), this is negligible. Enterprise deployments requiring 50+ concurrent agents will need horizontal scaling of the state server using Redis for session storage.
Ecosystem & Alternatives
Integration Points
Paseo leverages the emerging ADE (Agent Development Environment) pattern, positioning itself as the control layer rather than the agent itself:
- Claude Code: Full adapter supporting tool-use and extended thinking modes
- OpenAI Codex: Real-time streaming via CLI bridge with cost tracking
- Gemini CLI: Experimental adapter for Google's agentic tools
- Git Providers: Native webhooks for PR review workflows; GitHub/GitLab OAuth for mobile approvals
Adoption & Extensibility
Despite being approximately 3-4 weeks old (indicated by 0% 30-day velocity baseline), Paseo has gained traction among distributed teams using it for "agent babysitting"—maintaining oversight of long-running coding tasks from mobile devices while away from workstations.
The TypeScript codebase exposes a plugin API for custom agent adapters, suggesting a path toward community contributions for niche agents (Pi, local LLMs via Ollama, MCP tools). However, the ecosystem currently lacks enterprise-critical features: SSO/SAML integration, audit logging for compliance, and RBAC for agent permissions. Until these ship, adoption will remain limited to individual developers and permissive startups.
Momentum Analysis
AISignal exclusive — based on live signal data
| Metric | Value | Interpretation |
|---|---|---|
| Weekly Growth | +76 stars/week | Sustained viral interest in AI agent mobility |
| 7-day Velocity | 67.7% | Recent HN/Product Hunt feature or major release driving acceleration |
| 30-day Velocity | 0.0% | Project is <4 weeks old; establishing baseline |
| Fork Ratio | 13.8:1 (stars:forks) | High curiosity, low contribution (typical for early tools) |
Adoption Phase Analysis
Paseo is in early breakout—functionality exists but APIs are likely unstable. The 1,080 stars with only 78 forks suggests passive interest rather than deep community investment. The "orchestration" keyword in AI tooling is white-hot, and Paseo's mobile-first differentiation protects it from the 50+ generic "AI coding wrapper" projects launched this quarter.
Forward-Looking Assessment
The project addresses a genuine pain point that will intensify as coding agents run longer autonomously. Key risks: (1) Incumbent agents (Claude Code, Codex) adding native cloud/sync features, obviating third-party orchestration, and (2) security concerns around exposing coding agents to mobile attack surfaces. If the team ships enterprise security features (audit logs, RBAC) and a stable plugin API within 60 days, this could become the de facto "remote control" standard. Otherwise, it risks becoming a feature acquisition target or fading as native tools close the mobile gap.