They just haven't told you yet.

More than half of generative AI users already use AI without telling their employers, and 78% of knowledge workers bring their own AI tools to work. Your workforce is already building shadow workflows, stitching together AI agents, automating procurement shortcuts, and inventing cross-functional pipelines that do not appear on any org chart. The question is no longer whether the coordination function is shifting. It is whether you will surface, govern, and compound what people have already discovered.

The personal agent revolution has a governance problem. As information routing becomes computationally cheap, the strategic challenge shifts from moving information around the organisation to deciding where AI can be trusted, where human judgment must remain active, and how shared meaning is maintained across teams and agents. Tools like Claude Code have shown that autonomous AI agents are powerful, popular, and ready to act. They've also shown what happens when agents operate without shared semantics, formal reasoning, or organisational guardrails: unauthorised actions, prompt injection attacks, and enterprises deploying security scanners just to detect rogue agent instances on their own networks.

VisionClaw takes the opposite approach. Governance isn't an inhibitor, it's an accelerant — a way to turn shadow workflows into auditable, reusable organisational capability.

Every organisation rots in the same three directions.

Tribal knowledge accumulates in personal notes, Slack threads, half-finished docs — rich and current but illegible to agents, query tools, or anyone who wasn't in the room.

Formal ontology stales — expensive to author, expensive to keep clean, and nobody uses it.

Individual AI productivity does not compound — a skill a contributor develops on Monday is invisible on Tuesday; a routine they stabilise is a private shell alias; an insight they harvest from Pod memory is trapped in their screen.

Agents sit in the middle, forced to improvise shared meaning on every request. They hallucinate categories, fabricate relationships, route on keyword overlap.

VisionClaw treats all three as one problem with one mechanism: let the work become the institution — under human governance, visibly, as the graph itself.

The Contributor Stratum is the compounding loop between the other two. Private work lives in a contributor's Pod; a governed ShareIntent promotes it to Team, then through Broker review to Mesh, where it becomes institutional capability. The same mechanism handles notes, skills, workflows, ontology candidates, and reusable graph views.

A workspace, not a chat window. The /studio surface is a peer of /broker that puts ontology context, AI partners, installed skills, and share controls side-by-side with the work itself. Four pillars:

• 🪟 Sovereign Workspace — NIP-07 auth propagates transparently to Solid Pod and MCP; every artefact lands in /private/ by default; graph-node selection deep-links into the AI partner lane.
• 🥋 Mesh Dojo — decentralised skill sharing via publicTypeIndex.jsonld (ADR-029); SKILL.md files live in /public/skills/{slug}/ and federate to peer pods; eval suites gate team and mesh promotion (Anthropic Skill-Creator v2 discipline).
• 🌿 Ontology Sensei — background synthesis over /private/agent-memory/episodic/; proactive "3 skills you might run" nudges scoped to the current workspace focus; accept/dismiss feeds BC15 hit-rate KPI.
• ⏰ Pod-Native Automations — cron routines in /private/automations/*.json run via NIP-26 scoped delegation caps; output lands in /inbox/{agent-ns}/; offline routines can't cross to Mesh without broker review.

Three share states, strictly monotonic: Private → Team → Mesh. The only way down is ContributorRevocation or BrokerRevocation; every transition is a Policy Engine (ADR-045) decision with an append-only audit entry in /private/contributor-profile/share-log.jsonld.

Documentation: PRD-003 · ADR-057 · BC18/BC19 DDD · Sprint master

Every capability in the substrate serves this one event. A public:: true Logseq note accumulates signals — ontology wikilinks, agent proposals, authoring maturity — until it surfaces in a Judgment Broker's inbox. The broker takes two minutes, approves, the system opens a GitHub PR, and on merge the note becomes a live OWL class Whelk can reason over. The 3D physics view redraws around it. A Nostr-signed provenance bead closes the chain.

The ontology grows from the bottom up, not the top down. Nothing formal is authored unless something informal earned it.

KGNode (fast-moving notes) and OntologyClass (slow-moving canon) share one canonical IRI scheme vc:{domain}/{slug}. They connect through BRIDGE_TO edges that advance candidate → promoted via the migration event — or stay colocated when a note and class are the same concept from two angles. See ADR-048.

No LLM opinion-as-fact. Eight structural signals feed a sigmoid. Above 0.60 a candidate surfaces to the broker. Below 0.35 for three days it auto-expires. Agent confidence is one signal of eight — it does not bypass the gate. See candidate-scoring.md.

Shadow workflows become governed patterns. Shadow concepts become governed classes. Same five-beat loop, different units. See the Insight Migration Loop master design.

The CUDA physics is not decoration. It reads owl:physicality, owl:role, and maturity properties and turns them into five semantic forces: abstract clumps with abstract, processes band together, authoritative nodes anchor a stable core while drafts drift outward, orphaned notes (no ontology anchor yet) repel into a distinct zone that makes them visible as migration candidates.

92 CUDA kernel functions. 55× over single-threaded CPU. Every force is tunable per-domain. See physics mapping.

Every axis of this tool has good prior art. Obsidian and Logseq do wikilinks. Protégé and TopBraid do formal OWL authoring. Palantir does enterprise ontology. NotebookLM does LLM over notebooks. Neo4j Bloom does graph interaction. Ramp Glass does institutional AI workspace.

Nobody puts them in the same room. The upper-right quadrant — high formality and high interactivity and institutional compounding — was empty because the migration event itself is the hard part, and because the compounding funnel from private work to mesh asset was nobody's product surface. VisionClaw lives there. See prior-art analysis and the Contributor Stratum evidence annex for how the industry (PwC, McKinsey, a16z, Ramp, Anthropic) arrived at the same diagnosis from a different direction.

git clone https://github.com/DreamLab-AI/VisionClaw.git
cd VisionClaw && cp .env.example .env
docker-compose --profile dev up -d

VisionClaw is a hexagonal Rust backend with 21+ Actix actors, 9 ports and 12 adapters, 114+ CQRS handlers, CUDA 13.1 compute, and OWL 2 EL reasoning via Whelk-rs.

Every wikilink target becomes a first-class :KGNode regardless of publish state. Public nodes appear with full label and metadata; private nodes appear as topology-only (node shape and edges visible, label and metadata opacified via bit 29 on node_id in the V5 binary protocol and stripped from REST).

Each user's content lives in their own Solid Pod: /public/kg/ for published pages (world-readable, canonical URI), /private/kg/ for working graph (owner-only), and the full Contributor Stratum container set for their Studio state, automations, skills, and inbox. The Pod is write-master; the backend serves as indexer, aggregation point, and physics engine. The system never writes outside the owner's container.

VisionClaw aligns with two external ecosystems.

URN-Solid registry — We emit owl:sameAs urn:solid:<Name> on :OntologyClass entries where well-known vocabulary equivalents exist (Person, Document, Event). Each user's Pod publishes /public/kg/corpus.jsonl — a line-delimited JSON-LD snapshot following the URN-Solid registry generation convention. See the URN-Solid registry.

solid-schema — JSON Schema 2020-12 contracts for urn:solid: types, sitting between vocabulary (URN-Solid) and runtime (LOSOS). We publish /public/schema/kg-node.schema.json following the solid-schema convention so the same contract can be submitted upstream for ecosystem-wide adoption. solid-pod-rs validates JSON-LD PUTs against user-published schemas via the jsonschema Rust crate. See the solid-schema registry.

Solid-Apps (LOSOS) — We publish /public/schema/kg-node.schema.json and /public/schema/manifest.jsonld with the urn:solid:KGNode type binding so LOSOS apps built on LION + solid-schema + solid-panes + LOSOS can render any user's KG content directly from their Pod without VisionClaw-specific code. See the Solid-Apps project.

This alignment is behind feature flag URN_SOLID_ALIGNMENT=true. Full specification: ADR-054.

Key architecture decisions: ADR-040 · ADR-041 · ADR-042 · ADR-043 · ADR-044 · ADR-045 · ADR-046 · ADR-047 · ADR-048 · ADR-049 · ADR-057

Sovereign-mesh ADRs (Wave 1–4): ADR-028-ext · ADR-029 · ADR-030 · ADR-030-ext · ADR-050 · ADR-051 · ADR-052 · ADR-053 · ADR-054 · ADR-055 · ADR-056

VisionClaw uses the Diataxis framework — 144+ markdown files across four categories, 54+ with embedded Mermaid diagrams.

Entry points:

• Full Documentation Hub
• Contributor AI Support Stratum — sprint master · PRD-003 · ADR-057 · BC18/BC19 DDD
• Insight Migration Loop — master design · PRD · explanation · tutorial
• System Overview · Architecture Self-Review
• Deployment Guide · Quest 3 VR Setup
• Agent Orchestration · REST API · WebSocket V5 Binary

See the Contributing Guide.

Mozilla Public License 2.0 — Use commercially, modify freely, share changes to MPL files.