Think about the product. Let agents build it.
Most AI development tools start from code and build up — adding chat panels to IDEs, wrapping terminals in dashboards, bolting project management onto engineering workflows. The user they imagine is an engineer who wants AI help writing code faster.
Sidespace starts from the opposite direction. It builds down from product thinking.
The person orchestrating AI-assisted development isn't always an engineer. They might be a product designer, a PM, or a founder who thinks in terms of themes, features, priorities, and sequences — not file trees and build systems. Sidespace gives that person familiar tools to organize their thinking at a product level, and then connects that thinking directly to agents that can execute it.
The terminal is the engine room. The product surface is the bridge.
Sidespace's workspace tools — themes, roadmap, ideas funnel, kanban, notes, documents, whiteboards — are the primary interface. They aren't project management bolted onto a terminal. They are how the user communicates intent to agents.
A feature on the roadmap tells agents what's being built and when. A design doc linked to that feature gives agents the spec without asking. Notes capture informal thinking that agents can search. The ideas funnel signals what the user is considering. Task assignments tell agents what to work on. Every surface is a communication channel between the human and the agents, and the richest context comes from accumulated product artifacts — not a single chat conversation.
Every agent — Hoshi (the embedded assistant), Claude Code (in terminals), Umbra (the research agent), Kosmos (the maintenance agent) — reads from and writes to the same memory bank. The memory doesn't belong to any one agent or any one session. It is an independent organ that accumulates context over time.
This means a decision made during a terminal session is available to Hoshi in the chat panel. Learnings from one project can inform work on another. Starting a new session means picking up where the last one left off. The knowledge system currently holds over 1,500 memories, with automated research continuously feeding in new context and automated grooming keeping the signal-to-noise ratio high.
Squad View transforms the terminal into a coordination surface. Multiple CLI agents run simultaneously, each with a constellation name (Orion, Lyra, Vela) for identity across sessions and the memory system. The human acts as quarterback — assigning work, monitoring progress, making coordination decisions — while agents share context through the same tool surface and memory layer.
Three layers, each with a single responsibility:
| Layer | Platform | Role |
|---|---|---|
| Data | Supabase | Persistent storage, auth, real-time subscriptions |
| Compute | Railway | AI brain, 54 MCP tools, agent pipelines, scheduling |
| Workspace | Tauri | Native desktop shell — terminals, filesystem, git, UI |
All agents share a unified tool surface through the Sidespace MCP server — 54 tools for reading and writing projects, tasks, memories, and documents. A task created by Claude Code in a terminal is immediately visible to Hoshi in the chat panel. Research findings from Umbra surface in project context. The tool surface is agent-agnostic and portable.
| Agent | Role | Interface |
|---|---|---|
| Hoshi | Primary assistant and orchestrator. Project management, tool execution, context synthesis, diagram generation. | In-app chat panel |
| Claude Code | Hands-on coding agent. Writes, edits, and debugs code in live terminal sessions. | Squad View terminals |
| Umbra | Research and discovery. Scans external sources, stages findings for quality review. | Background pipeline |
| Kosmos | Memory operations. Grooming, deduplication, healing, transcript mining. | Scheduled pipelines |
| Atlas | Documentation agent. Keeps docs aligned with the codebase. | GitHub Actions |
Sidespace v0.1.0 is in active development and serves as its own primary development environment. The product surface, memory organ, agent ecosystem, and multi-agent cockpit are all operational.