OrgEdge
SysEdge — open-source ontology for multi-agent development

Twelve Claude sessions. One queryable ontology.

Give yourself the edge. In a knowledge graph, an edge is a relationship with meaning. SysEdge builds them across your entire codebase — stories realised by use cases, features verified by tests, decisions addressing standards — giving every parallel Claude Code session the edge of instant orientation over sessions that start from scratch. Measured: 96% fewer orientation tokens, four tasks, one real codebase.

The coordination problem

Twelve sessions. One codebase. Zero shared memory.

Without SysEdge

  • Each session re-reads the same source files to figure out what exists — 20–30% of every context window spent re-orienting
  • Two sessions independently start work on the same feature. One gets merged, the other discarded.
  • A test is added but nobody knows which feature it covers. Coverage looks green; the gap is invisible.
  • An architectural decision is made in one session. The other eleven never see it.
  • "What's left to do?" requires a human to reconcile eleven CONTEXT.md files.

With SysEdge

  • Sessions query the graph: 30 seconds to know every feature, module, endpoint, and open task
  • start-enhancement flips a node to in-progress. Other sessions see it and route to different work.
  • Every test is linked by VERIFIES edges to the symbol, feature, and use case it covers — gaps surface as graph queries
  • ADRs are nodes with ADDRESSES edges to 53 standards. Every session sees the same architecture state.
  • worklog returns a prioritised queue per session. No human reconciliation required.
The model

Full requirements traceability.

Every artefact from user story to unit test lives as a node. The graph is the specification.

User Story → REALIZED_BY → Use Case → REQUIRES → Feature → PROVIDED_BY → Module → CONTAINS_SYMBOL → Symbol / Endpoint
VERIFIES →   Test (4 tiers)

A gap anywhere in this chain is visible instantly. There is no "I thought that was tested" — the graph either has the VERIFIES edge or it doesn't.

The knowledge graph

Why an ontology, not just a database?

A task tracker stores rows. An ontology stores meaning. In SysEdge, a Feature is not a ticket — it is a capability a Module provides. A UseCase is not a description — it is a user-visible flow that realises a Story. A Test's V-model tier is derived from which spec artefact it verifies, not from a tag. The 53 architecture standards are nodes with ADDRESSES edges to ADRs; a gap is a standard with no incoming edge. This is what makes the graph answer questions a flat database cannot without bespoke code.

Defined semantics matter most when twelve sessions share the graph simultaneously. "Which features have no use case?" or "which standards have no addressing decision?" are answered in one query — because the relationships encode what the words mean, not just how they were entered. Without an ontology, every session re-derives these answers from prose. With one, the graph derives them once and serves all sessions equally.

$ # Which standards have no Architecture Decision Record? $ sys_graph.py analyse --orphans --instance architect ADRs without addressed standards (3): SEC-BP02 Audit log for write operations → no ADR OPS-S03 Metrics wired to all service binaries → no ADR INFRA-BP01 Container images pinned to versions → no ADR $ # Which features have no use case covering them? $ sys_graph.py analyse --orphans --instance plan Features without UC (4): F-FNS-007 Access path to function planner F-IMP-003 Plan versioning with SUPERSEDES edges F-CMP-002 Compliance recheck endpoint F-GAP-003 Idea analyser
Session start in 30 seconds

Orient without reading source.

Three commands replace ten minutes of file reading.

$ python3 scripts/sys_graph.py briefing --instance manage 2>/dev/null ══════════════════════════════════════════════════════════ SYSTEM GRAPH BRIEFING — manage (2026-05-22) ══════════════════════════════════════════════════════════ MOD-delegation Delegation 12/12 ✓ ✓cmp 12/12 ✓int 12/12 ~uc 8/12 ✗e2e 0/12 MOD-succession Succession 9/9 ✓ ✓cmp 9/9 ✓int 9/9 ✓uc 9/9 ✗e2e 0/9 MOD-positions Positions 11/11 ✓ ✓cmp 11/11 ✓int 11/11 ✗uc 0/11 ✗e2e 0/11 ENHANCEMENTS (3) PROPOSED (3) ENH-330 [Must] Delegation inbox API endpoint → F-DLG-003 ENH-331 [Should] Bulk reassign on position offboarding → F-DLG-008 ENH-354 [Could] Add dependency health checks to /health $ python3 scripts/sys_graph.py worklog --instance manage 2>/dev/null ★★★ ENH-330 [Must] Delegation inbox API endpoint — GET /manage/delegation/inbox Feature : F-DLG-003 Delegation inbox view Module : MOD-delegation backend/internal/p8/delegation/ $ python3 scripts/sys_graph.py start-enhancement --id ENH-330 --instance manage ✓ ENH-330 in-progress — Delegation inbox API endpoint Other sessions can now see this is being built.

Fig. — SysEdge in action: the web visualiser navigating a real Go/TypeScript codebase (PocketBase), followed by the CLI demo showing briefing, worklog, stories, test-gaps, and export in under 60 seconds.

Test architecture

V-model coverage, enforced by the graph.

SysEdge implements the V-model: each specification artefact has a required verification artefact at the same abstraction level. Four tiers, each visible in the session briefing.

Specification Test tier Technology Owned by
User Story e2e Playwright — full cross-tool journey master session
Use Case usecase Playwright — single UC flow owning session
Feature / Module integration Python pytest — API contract owning session
Symbol / Routine component Go test, vitest — function level owning session

Gaps at any tier are surfaced in test-gaps --instance X with the exact test file to create. No manual audit required.

Architecture governance

Standards as graph nodes. Compliance as edges.

53 industry standards across Execution, Development, Operations, Security, and Infrastructure architecture are stored in the graph, sourced from 12-factor, OWASP, NIST, SRE practice, and cloud-native patterns.

Each standard links to an Architecture Decision Record (ADR) documenting how your codebase addresses it via an ADDRESSES edge. Standards with no addressing ADR are gaps — returned by a single query. An initial audit on a live codebase identified 14 genuine gaps and 8 compliant-but-undocumented standards in one session.

$ # Standards with no addressing decision = architecture gaps [SECURITY] SEC-BP02 Audit log for write operations → no ADR SEC-BP03 Rate limiting on auth endpoints → no ADR [OPERATIONS] OPS-S03 Metrics wired to all service binaries → no ADR OPS-BP03 Runbooks for Sev-1 incidents → no ADR
Import · Export · Analyse

Your existing documents become graph nodes.

Import requirements documents, architecture specs, and use case files directly into the graph. Export the graph back as structured Markdown. Analyse for duplicate features, split candidates, and orphan nodes — all through Claude Code skills.

IMPORT SKILLS

  • /import-stories — User story documents → SysUserStory nodes
  • /import-use-cases — UC documents → SysUseCase nodes + flows
  • /import-requirements — Feature lists → SysFeature nodes
  • /import-architecture — Arch docs → ADRs + standards
  • /import-code — Existing symbols → inferred feature clusters

EXPORT & ANALYSE

  • /export --type stories — User stories as Markdown
  • /export --type application-arch — Module/feature map
  • /export --type technical-arch — ADRs + standards
  • /export --type test-coverage — V-model coverage report
  • /analyse-graph — Merge candidates, split candidates, orphans

Every import uses a review-and-confirm step — no silent writes. Near-duplicates are flagged before commit.

Bootstrap kit

Zero-config setup via Claude Code scan.

The biggest adoption barrier is the initial seed JSON — defining your modules, features, and test packages. The /init-sysgraph skill solves this in one command. Included in the bootstrap kit.

$ # Inside a Claude Code session, from your working directory: /init-sysgraph Scanning working directory... Detected: Go backend (backend/internal/) — 12 packages Detected: TypeScript frontend (src/features/) — 8 feature modules Detected: Python tests (tests/) — 47 test files Detected: Existing modules → 0 (new project) Generating seed configuration... ✓ 20 SysModule nodes derived from package structure ✓ 89 SysFeature nodes derived from exported symbols ✓ 3 SysTestPackage groups (component / integration / e2e) ✓ seed written to data/sys-init-YYYY-MM-DD.json Next: python3 scripts/sys_graph.py seed data/sys-init-YYYY-MM-DD.json

Claude Code reads your directory structure, infers module boundaries from package layout, derives features from exported symbols, and generates a complete seed file — ready to load into Neo4j in under two minutes.

Measured on a real codebase

96% fewer orientation tokens.
Pays for itself in under a week.

We measured the tokens a Claude Code session reads before writing the first line of code — the orientation phase. Four tasks across a live 44-module Go/TypeScript codebase: two defects, two enhancements, ranging from a small service to a 2,500-line cross-layer feature.

Task Module size Without SysEdge With SysEdge Saving
DEF-001 — MCP scope enforcement defect
licmcp · small service · 1 file
small 21,929 1,348 −93%
ENH-103 — Remove legacy impl-planner surface
plan · 4 Go files + 1 TSX component
medium 31,467 1,683 −94%
ENH-330 — Delegation inbox API endpoint
manage · 2,889-line handler file
large 48,588 1,423 −97%
ENH-200 — CreateActionDialog + reassign endpoint
core · cross-layer: Go handler + TSX + API
large 53,114 1,086 −97%
Average 38,774 1,385 −96%

AT 10 SESSIONS/DAY · SONNET 4.6

Without SysEdge $31.73/month
With SysEdge $1.25/month

orientation cost only · implementation tokens are equal

PAYBACK PERIOD · $149 BOOTSTRAP KIT

5 days

at 10 sessions/day. Faster at higher session counts.

token savings alone · coordination and coverage benefits not included

How we measured this ↓

We measured the orientation phase only — the tokens a session reads before writing any code. Implementation tokens (reading source to understand what to change, writing the fix) are equal in both paths and excluded from the comparison.

Without SysEdge: we listed every file a Claude Code session would realistically read to orient on the task — the root CLAUDE.md, the instance CLAUDE.md, the enhancements or defects file, and the relevant source files. File sizes are actual bytes from the live codebase.

With SysEdge: we captured the actual output of briefing --instance X, worklog --instance X, and show-enhancement --id ENH-NNN for each task, plus the instance CLAUDE.md (read in both scenarios).

Token estimate: len(text) / 4 — a conservative approximation for mixed code and prose. Claude's actual tokeniser (BPE) typically gives slightly fewer tokens for code-heavy content, so the real saving may be larger.

Caveat: the "without" scenario reads whole source files. A skilled session using Read with offset/limit on targeted sections would read fewer tokens — we estimate the honest real-world saving at 60–80% for medium and large modules.

All four tasks are real completed work on a live 44-module Go/TypeScript codebase. The codebase, measurement script, and raw outputs are included in the bootstrap kit.

Test coverage

Coverage as a graph query — not a percentage.

Coverage isn't one number; it's a shape across four V-model tiers. SysEdge maps every test to the spec artefact it verifies — e2e tests to user stories, Playwright flows to use cases, integration tests to features, unit tests to symbols. After one day of graph-guided sessions on a 12-instance system, 9 of 11 instances reached 100% integration coverage. 1,459 vitest tests were auto-scanned and 5,807 VERIFIES edges created with no manual mapping.

COVERAGE AFTER ONE DAY — 196 features across 9 active instances

InstanceFeaturescmpintuc
core40/40 ✓54020
framework30/30 ✓213011
manage35/35 ✓353535
plan43/43 ✓284342
platform13/13 ✓131313
p1, licmcp, training, master35/35 ✓varies35varies
Total196/196112196121

cmp=component/unit · int=integration/API · uc=UI flow per use case
5,807 VERIFIES edges · 7,470 tests · 98 endpoints linked

WHAT SESSIONS REPORTED

"Without the graph I would have written a new unit test and felt done. Now I can see the shape — integration tests validate API contracts but do not test component rendering or specific user flows."
— core session, FB-083

"I wrote tests specifically for 6 features with zero component coverage even though I hadn't touched those components directly."
— framework session, FB-063

"MOD-board 0/4 cmp makes visible a gap I would otherwise rationalise away — integration tests exist so it feels covered."
— plan session, FB-068

Language support

Language-agnostic graph. First-class scan for Go, TypeScript, Python, Java, and C#.

The graph, worklog, briefing, test coverage tracking, architecture standards, and all coordination features work for any language. Code symbol extraction and test scanning add richer navigational detail for supported languages.

Go
Full scan: functions, handlers, test files
TypeScript
Full scan: components, hooks, exports
Python
Full scan: classes, functions, pytest files
Java
Full scan: public classes and methods
C#
Full scan: public classes and methods
Rust / other
Graph features work · manual symbol linking

Adding a new language scanner is one Python function (~20 lines). See INSTALL.md for the plugin interface.

How it works

Neo4j + a 3,000-line CLI + a skill file.

Components

  • Neo4j Community (free) — the graph database
  • sys_graph.py — CLI for all graph operations, 40+ commands
  • sys_graph_viz.py — local web visualiser, drill-down navigation
  • feature-graph skill — Claude Code skill file, governs all session behaviour
  • Seed JSON — your project's module and feature definitions

Safety model

  • Sessions use MERGE (never DELETE)
  • Graph reset is a separate shell script — not accessible from Claude
  • Instance-filtered seed auto-backs up before running
  • VERIFIES edges snapshot before any seed — post-backup links preserved
  • Sessions only write nodes within their declared instance scope
Pricing

Free CLI. Pay for the visualiser and bootstrap kit.

The CLI gives you full graph power from the terminal. The bootstrap kit adds the web visualiser, Docker setup, auto-seed skill, and session templates — everything to be productive in an afternoon.

Feature Free CLI Bootstrap Kit · $149
FOUNDATION
Briefing, worklog, test-gaps per session
Scan code + tests (Go, TypeScript, Python, Java, C#)
Create + track enhancements, defects, user stories
V-model 4-tier coverage (component / integration / UC / e2e)
Backup + per-instance restore
MIT + Commons Clause licence + README + INSTANCES guide
BOOTSTRAP KIT ONLY
sys_graph_viz.py — web visualiser
drill-down graph, entity editor, coverage tiers, export panel
Export commands — stories, use-cases, application-arch, technical-arch, test-coverage
Import skills — /import-stories, /import-use-cases, /import-requirements, /import-architecture
Graph analysis — merge candidates, split candidates, orphan detection
/init-sysgraph — auto-seed from working directory
Architecture standards YAML (53 standards, 5 domains)
Docker Compose + one-command setup (python3 setup.py)
All Claude Code skill files + session role templates
12 months of updates + email support
Get the free CLI on GitHub
Buy bootstrap kit — $149 →

Per repository · one-time · instant download · VAT included

20+ parallel sessions or custom language support — contact us.

The free CLI is MIT + Commons Clause — free to use for your own projects, including commercial software.
You may not sell the CLI itself as a product or service without a commercial licence.
The bootstrap kit is licensed per repository with 12 months of updates and support.
One licence covers unlimited team members and machines working on that repository.
Read the full licence conditions →

Installation

Running in four steps.

# 1. Start Neo4j (requires Docker) docker run -d --name sysgraph-neo4j \ -e NEO4J_AUTH=neo4j/yourpassword \ -p 7687:7687 -p 7474:7474 \ neo4j:5.26-community # 2. Install Python dependencies pip install neo4j python-dotenv # 3. Configure connection echo "NEO4J_URI=bolt://localhost:7687 NEO4J_USER=neo4j NEO4J_PASSWORD=yourpassword" > .env # 4. Initialise constraints and load your project definition python3 scripts/sys_graph.py init python3 scripts/sys_graph.py seed data/your-project-seed.json ✓ Ready. Add the feature-graph skill to your Claude Code sessions.

The bootstrap kit includes a Docker Compose file that starts Neo4j with the right configuration and a setup wizard that generates your initial seed JSON from your working directory.

Validated on a 12-agent system

Stop re-reading source files. Query the graph.

SysEdge came out of running a 12-instance Claude Code system — twelve parallel sessions, one shared codebase, no human in the loop for routine coordination. The ontology is what keeps the sessions coherent: shared meaning, shared state, shared next-actions. When the graph is the source of truth, sessions spend their tokens building, not exploring.

96% fewer orientation tokens
measured on a real 44-module codebase
Zero duplicate implementations
start-enhancement marks work in-progress
V-model coverage in four tiers
component · integration · UC flow · E2E
53 architecture standards audited
gaps as graph queries · ADRs as edges
Import · export · analyse
documents in · structured reports out