Row-Bot 5.0 architecture overview

Architecture

How Row-Bot stays local-first.

These high-level subsystem maps show Row-Bot's 5.0 control plane: one React client, in the desktop window, a browser, or on a phone, talks to a local FastAPI server; one checkpointed parent coordinates agent work; conversations, documents, memory, workflows, and project assets stay durable with the host; and exact-model reasoning, context metering, and recoverable compaction remain provider-aware. Designs and code folders belong to the conversation that works on them, tools and skills can load progressively without expanding approval boundaries, OpenCode routes through native model protocols, every signed-in device shares one owner authority, and models, network routes, and optional runtimes remain explicit.

Durable local state

Conversations, orchestration checkpoints, documents, memory, tasks, access records, and project assets live with the Row-Bot host, served to one React client by a local FastAPI server.

Parent-led work

One authoritative parent joins required results, tracks dependencies and later waves, and finalizes once across retries, approvals, and restarts.

One owner authority

The desktop app, a browser on another computer, and a phone expose the same owner capability behind invitations, revocable device sessions, exact saved addresses, and host checks.

Explicit external routes

Hosted models, tools, channels, MCP, Computer Use, Tailscale, and reverse proxies remain separately configured, reviewable, and approval-aware.

Subsystem maps

Open any high-level diagram for a larger view. The gallery starts with the core-agent map, then moves through specialized subsystems; the technical reference documents the current 5.0 runtime and security contracts in detail.

Row-Bot core agent architecture diagram
01

Core Agent

Interfaces, the local server, parent-led orchestration, provider routing, checkpointed budgets, context, memory, tool calls, workflows, storage, and control planes.

Row-Bot context architecture diagram
02

Context Assembly

Complete next-input metering, model-aware limits, rolling compaction, durable summaries, and the current turns, skills, memories, documents, and tool results that become model-ready context.

Row-Bot knowledge graph architecture diagram
03

Knowledge Graph

Durable document queues, resumable extraction, atomic sharded retrieval, graph storage, wiki export, and Dream Cycle refinement.

Row-Bot background workflow architecture diagram
04

Background Workflows

Natural-language workflow creation, scheduling, execution, approvals, delivery, and recurring loops.

Row-Bot multi-channel architecture diagram
05

Multi-Channel Runtime

Messaging entry points, adapters, normalization, identity mapping, attachment handling, and delivery.

Row-Bot Design panel architecture diagram
06

Design Panel

Conversation-first creation, design kinds and brand, targeted edits and review in the Design panel, preview sandboxing, and present, share, and export paths.

Row-Bot Developer panel architecture diagram
07

Developer Panel

Code folders bound from the conversation, approval modes, changes, checks, and Git in the Developer panel, the terminal dock, Docker Sandbox, and Custom Tools.

Row-Bot skills system architecture diagram
08

Skills System

Smart Skills activation, task-scoped progressive selection, child isolation, Skills Hub imports, bundled guides, slash commands, and bounded skill prompts.

Row-Bot safety privacy and control architecture diagram
09

Safety, Privacy & Control

Local-first boundaries, owner authentication, exact trusted addresses, managed route ownership, approvals, prompt-injection defense, target-window control, encrypted credentials, and updates.

Row-Bot self-evolution architecture diagram
10

Self-Evolution

Status introspection, diagnostics, capability registry, insights, skill guidance, and controlled self-management.