Platform Overview
SystemDox sits between your architecture knowledge and your AI coding tools. It indexes requirements, ADRs, checks, and domain knowledge, then delivers relevant context through two channels: MCP Server (real-time) and Static Export (CLAUDE.md, .cursorrules).
MCP Server — Real-time Context
The MCP (Model Context Protocol) server is the highest-value integration. AI coding tools connect directly and query for context while the developer is working. The agent gets architecture decisions, checks, specs, and domain knowledge scoped to exactly what it’s building.
Available MCP Tools
get_architecture_decisions
Query ADRs by domain. Returns relevant architecture constraints.
get_checks
Query banned and required patterns by language and area.
get_specs
Retrieve Given/When/Then specifications for a feature.
search_knowledge
Free-text search across all indexed documentation.
Static Export — Context Files in Every Repo
For teams that aren’t ready for MCP, or as a baseline layer, SystemDox generates repo-specific context files. These are the same files you’d manually maintain — but kept current automatically from your single source of truth.
Generated Files
CLAUDE.md
For Claude Code
.cursorrules
For Cursor
copilot-instructions.md
For GitHub Copilot
Document Types That Matter for AI
Not all documentation is useful for AI agents. SystemDox focuses on the six document types that directly impact code generation quality.
| Type | Purpose | Example |
|---|---|---|
| ADRs | Architecture constraints | ADR-007 (PwebAuthoriser) |
| Checks | Banned/required patterns | ”Never use Sentry.captureException directly” |
| Specs | Feature behaviour (Given/When/Then) | shared-specs convention |
| Domain glossary | Term definitions | ”Principal = user_id + tenant_id + role + scopes” |
| Integration contracts | API shapes, events | Cross-repo dependency map |
| Runbooks | How to deploy/debug | ”CI_ENABLED variable controls pipeline” |
Bidirectional Sync
Unlike static documentation, SystemDox maintains a live, bidirectional sync. AI agents read context and write back discoveries. GitHub commits trigger re-indexing. CLAUDE.md files regenerate when architecture changes. Documentation stays current because the platform maintains it.
The Closed Loop
The loop runs across the three surfaces rather than beside them. You capture a decision and it becomes a check. Build delivers that check to the agent, the agent writes code against it, and architecture fitness tests verify the result in CI — validation lives inside Build, not in a stage of its own. What survives review is published to Knowledge Base, where your team and your AI agents both read it.
And when something does slip through, the violation goes back to Capture as a new check with a fitness test attached, so the same mistake cannot recur. That feedback edge is the learning step — a loop closing over all three surfaces, not a fourth one.
What Makes This Different
Bidirectional Sync
AI agents don’t just read — they write back. When Claude Code creates an ADR or discovers a pattern, it pushes it to SystemDox. Documentation stays current because the AI maintains it.
Context Ranking
Not all docs are relevant to every task. SystemDox scores and filters context based on what files the agent is touching, what domain it’s working in, and token budget constraints.
Check Enforcement
Instead of hoping developers read the docs, SystemDox feeds checks directly to the AI. The agent literally cannot ignore “don’t use CognitoAuthorizer” because it’s in its active context.
Living Specs
Specs in SystemDox become test generation inputs. Given/When/Then specs aren’t just documentation — they’re executable context that generates tests automatically.