Ground Your AI.
The sovereign memory layer for AI agents. Every document, drive and agent session - fused into one governed knowledge graph on your own infrastructure. Unlimited tokens. No vendor lock-in. Run locally, own your data.
curl -fsSL https://gctrl.tech/install | bashWorks with the AI you already use
Claude Code
GitHub Copilot
LangGraphYour AI already speaks GCTRL - connect over MCP with one config block. See all integrations →
The real problem
Your AI is only as good as your data - and most data isn’t ready.
Garbage in, garbage out. Before AI can deliver, your knowledge has to be accessible, clean, and governed - and in most enterprises it’s none of the three.
Locked in silos & legacy systems
The knowledge your AI needs is scattered across SharePoint, old mailservers, legacy SQL, and orphaned file shares. Before anything is useful, you have to reach it - and just reaching it is a project of its own.
Messy, duplicated, contradictory
Garbage in, garbage out. Point an LLM at raw, unresolved data and it learns from noise: the same entity under ten names, stale records, conflicting versions. The answers look confident and are quietly wrong.
Governance is overwhelming
Even once you can reach the data - who is allowed to see what? Classification, clearance, and an audit trail across every source is a task most teams start, dread, and never finish.
GCTRL exists to fix exactly this: it makes the messy, scattered, sensitive data your AI needs accessible, clean, and governed - so what goes in is worth what comes out.
How It Works
From raw sources to ground control.
One pipeline turns scattered, conflicting data into a governed memory your agents can actually trust.
Ingest
Connect any source - SharePoint, Google Drive, email archives, databases, APIs. A governed ingestion layer classifies and tags everything on the way in.
Resolve & fuse
FUSE collapses duplicates and contradictions across every source into one clean, canonical knowledge graph - no conflicting copies, one version of the truth.
Organise into memory
Facts land in layered memory - hot dossiers, warm chunks, the cold graph, and a curated Wiki - on top of your swappable Neo4j + Qdrant.
Serve to agents
Your agents query it over MCP - clearance-filtered and fully audited - returning grounded answers with provenance, not guesses.
One governed flow - ingestion to answer - with provenance preserved at every hop.
Explore the architectureHow it feels
From scattered noise to one source of truth.
Team Chat 1
Team members contribute knowledge straight from their chats - into private or shared compilations, at their clearance level.
Team Chat 2
Every conversation can become durable knowledge: decisions, facts and context are extracted, not lost in scrollback.
Team Chat 3
Private by default, shared when you choose - each compilation has its own access rules.
Agent Session 1
Claude, Codex & co. write what they learn in each session back to shared memory - the team’s agents stop forgetting.
Agent Session 2
Scoped tokens govern what each agent may read and write - an agent only ever sees what it’s cleared for.
Agent Session 3
Swap the agent, keep the memory: session knowledge lives in YOUR fabric, not inside the tool.
PDFs
PDFs are extracted into entities and relations - with provenance preserved for every fact.
Docs
Word, Excel and wiki pages become structured knowledge instead of dead files.
Legacy System
Old ERP exports, file shares and orphaned apps stream in through the governed ingestion layer.
Cloud Drive
Google Drive & SharePoint sync continuously - new files are extracted as they land.
SQL Database
Structured records join the same graph - finally connected to the unstructured world.
Mail Server
Mail archives become searchable knowledge - clearance and privacy intact.
Sources & Agents
Sources
SharePoint
Google Drive
Email & files
Agents
Claude
Codex
Hermes
GCTRL · Middleware
Ingestion
classify
Access rights
clearance
GCTRL
core
Memory layers
Hot
dossiers
Warm
chunks
Cold
graph
Wiki
pages
Your Infrastructure · swappable
Postgres
Qdrant
Neo4j
Wiki
Middleware that sits on top of whatever you already run
Sources flow in, agents plug in, and access control governs both - while GCTRL organises everything into layered memory on top of swappable storage. Bundled for a one-line install; point it at your own Neo4j, Qdrant and Postgres anytime. No lock-in.
Ingest from any source
SharePoint, Google Drive, email archives and other silos stream in through a governed ingestion layer.
Full classification control
Per-element clearance on nodes, edges and chunks. Scoped tokens for every user and agent. Merge everyone’s knowledge into one graph and classification still holds - each person sees only what they’re cleared for. Granular, auditable, on every read and write.
Parallel memory layers
Hot dossiers, warm chunks, the cold graph, and a curated Wiki - each backed by its own store, organised on a high-performance core.
Your agents plug in over MCP
Claude, Codex, Hermes and more gain durable, access-controlled memory as a team member.
Trust, built in
Running at the speed of trust.
Anyone can bolt an LLM onto a database. The hard, unglamorous part is controlling exactly who - human or agent - can touch which fact, and proving it forever after. We built GCTRL knowing that in the enterprise, compliance isn’t a feature - it’s the permission to exist.
Built for environments governed by these frameworks - design posture, not a certification claim. How we address each
Per-element classification
Nodes, edges, and chunks each carry their own clearance markings. Sensitivity travels with the data, not the schema - so it survives every merge, query, and export.
Scoped tokens for users AND agents
Issue narrow, time-bound, revocable capabilities to a person or an AI agent. Every retrieval is filtered server-side against the caller’s scope - never a client-side hint that can be ignored.
Forensic audit trail
Every access, every denial, every scope grant - captured with the caller, the context, and the verdict. The receipts your CISO, auditors, and DPO accept before procurement signs.
Granular orchestration
Merge the whole organisation’s knowledge into one graph - and classification still holds. Two people of different clearance query the same data and each sees only what they’re cleared for.
And it’s effectively free. Classification and scope enforcement add ≈ 0 ms to retrieval - security with no performance tax.
See the facts ↓Integrations
Connect your entire data estate.
Native connectors for the tools your teams already use. Replace the backend without changing the frontend.
Google DriveCloud document source
ConfluenceWiki & documentation
GitHubCode & docs
ObsidianKnowledge vault
Neo4jSwappable KG store
QdrantSwappable vector storePlus REST API, webhooks, and a connector SDK for any custom source.
A code brain that travels with you.
Point GCTRL at a repository and its structure becomes a knowledge graph - files, classes, functions, who-calls-what - living in your cloud, shared by every agent you use. Switch harness, switch machine, switch model: the knowledge comes along. And your agents spend their tokens on thinking, not on grep.
94%
Fewer tokens per structural question
GCTRL repo · Rust + TS + Python · 12/12 correct
90%
Fewer tokens on a 1,000-file TypeScript app
12/12 correct · one call per answer
100%
Call-graph edges verified correct
1,119 edges cross-checked against the compiler
sec
Re-index after a commit
incremental - only changed files travel
One code brain, every harness
Claude Code on your laptop, Cursor at the office, Codex in CI, an Anvil agent on the server - they all connect to the same Codebase KB. What one agent learned about your repository, the next one already knows. Your context stops living in a single chat window.
The graph answers, grep does not
Where is X defined, who calls it, what breaks if I change it - one call to the graph, read only the lines that matter. No more scanning whole trees into the context window and paying for it on every turn.
Decisions live next to the code
The why and the where in the same graph: architecture decisions, conventions and gotchas are stored on the very symbols they concern, so the next session inherits them instead of rediscovering them.
Yours, scoped, on your terms
Runs in your cloud or on-prem. Every token carries its own Codebase access; a colleague’s agent sees the repositories it was granted and nothing else. Indexing is a line you write - code never leaves the machine by accident.
Two lines to switch it on: create an access token, run gctrl init in your repository - the agent indexes it on start-up and follows the coding protocol from the GCTRL skill.
Near-supervised quality. Zero training.
Measured on standard public benchmarks and our own testbench - fully local, no labelled data.
0.97F1
Entity linking (DBLP-ACM)
0.978recall
NER detection (bilingual)
unsupervised · zero training
<50ms p95
Memory retrieval latency
7-27 ms median
≈0ms
Access-control overhead
compliance is effectively free
2,750/s
Matching-engine throughput
sub-quadratic ~O(n^1.5)
| Dataset | GCTRL (measured) | Published SOTA (cited) |
|---|---|---|
| DBLP-ACM | 0.97no training | Ditto 0.989 · DeepMatcher 0.985 |
| Abt-Buy | 0.866local embeddings | Ditto 0.891 · DeepMatcher 0.628 |
Supervised baselines are cited from their published papers (Ditto, DeepMatcher) on the identical public datasets - GCTRL reaches comparable quality with no labelled training data. Head-to-heads vs other GraphRAG systems are in progress.
Performance tip: vector search is the slowest retrieval step (~44 ms vs ~7 ms for graph). Qdrant is swappable - point GCTRL at a faster vector store to cut query latency further.
Full benchmarks →Your AI deserves
solid ground to stand on.
Stop building on unstructured noise. GCTRL gives your enterprise AI the knowledge foundation it needs to be accurate, explainable, and trustworthy.
1 · Install locally - free
One command, fully self-hosted. Unlimited tokens from day one.
2 · Connect your AI
Claude Code, Codex, Cursor or any MCP client - one config block.
3 · Scale with scoped tokens
One scoped access token per colleague, project or agent - each sees only what you grant it.
Fully on-premises · GDPR-ready · No vendor lock-in