ContextVerse

Context as a Service.

Every AI tool wants to own your memory. ContextVerse gives it back to you.

One context for every AI tool — Claude, ChatGPT, Cursor, Copilot. Install in 1 minute, configure in 2. Full control for your team or enterprise, with a single source of truth.

Self-hosted · Vendor-neutral · Open source

Documentation · Install · Quickstart · Templates

Status: early development. The commands, HTTP API and on-disk layout are not yet stable. Pin a release if you depend on them.

---

The problem

Claude has its projects. ChatGPT has its memory. Cursor has its rules. They don't talk to each other, they don't sync, and when you switch tools you start from zero.

For a team it is worse. There is no way to onboard a developer into your AI workflow, no way to share context across the team, no way to enforce AI rules company-wide. Every developer builds their own silo, and the knowledge lives in Slack messages nobody reads.

The solution

contextd is a single binary that holds one context space — plain Markdown you write and own — and delivers it into whatever entry point each AI tool actually reads.

You curate; the AI reads. Nothing is inferred, embedded, or stored behind your back.

contextd init                              # guided setup: solo, client, or server
cd ~/projects/api && contextd activate     # Claude, Cursor and Copilot now read the same space
contextd file history team/principles.md   # v4 — who changed the rules, and when

---

What it looks like in practice

Procedures your AI already has open

Your team's real knowledge isn't facts, it's procedures: which script to run, which one never to run, what to do when it fails. Today that lives in one senior engineer's head and gets re-explained to every AI, every session, by every developer.

Write it once, as a file you own:

# team/deploy.md

## Deploying to production

Production deploys go through ArgoCD — never run `helm upgrade` directly.

    ./scripts/deploy.sh <service> <version>     # deploy
    ./scripts/rollback.sh <service>             # rollback

If the sync hangs >5 min, page #platform-oncall before retrying.

After contextd activate, Claude, Cursor and Copilot are all handed that runbook — with your script paths — at the start of every session, instead of each improvising a plausible-looking helm command.

Honest ceiling: delivery is guaranteed, obedience is not. ContextVerse makes sure the model has your procedure; it cannot make a model obey it.

Context that can't rot silently

Stale context is worse than no context — an AI confidently repeating a deploy process you replaced six months ago is a bug you cannot see. So freshness is not left to someone remembering:

---
last-validated: 2026-07-26
stale-after: 30d
owner: platform-team
---

contextd freshness check          # what's stale, who owns it, since when
contextd freshness nag --server   # emit freshness.stale webhooks → Slack, Jira, PagerDuty
contextd freshness validate team/deploy.md

Your context becomes an auditable asset with an owner and a review cycle.

Onboarding, in one command

contextd init client --url https://context.my-team.com --token cv-alice-xxxx

The space is cloned, the developer's identity is set, and their AI tools are wired.

The same context on a second machine

identity/ ships as init-only: seeded once from the template, then it belongs to that machine and is never pushed back. That default protects a team space, where one person's me.md must not land on everyone. For your own two machines there is no team copy to protect, so pick whichever of these fits:

	How	When it fits
Rebuild	contextd init solo again, answer the questions	Two machines, different roles — a work laptop and a personal one
Git	contextd backend set git --url …	You already live in git and want the history to be the same history
Your own server	contextd init server, then point both machines at it	You want it automatic. Set contextd space sync set <space> identity/ always — on a personal server that is correct, on a shared one it is not

The tool does not decide this for you, because the right answer depends on whether the second machine is yours or someone else's, and only you know that.

---

What makes it different

Stated as our own capabilities, not as claims about anyone else's product:

Self-hosted on any infrastructure	Your server, your storage backend — filesystem, git, S3, SQL. No vendor account required to run it.
Path ACL, deny-by-default	Explicit deny wins, most-specific path wins. Access is a policy, not a sharing toggle.
Per-file version history	Every file has vN, a body, and an author. Restore any version.
Audit trail	Every operation records actor, timestamp and target, hash-chained.
Freshness with owners	Context declares when it expires and who owns it; staleness raises a webhook.
Vendor-neutral by construction	The same space is delivered to Claude, Cursor, Copilot and ChatGPT. Switching tools costs nothing.

---

Install

macOS	Windows
brew tap orkcom-tech/tap brew install orkcom-tech/tap/contextd	scoop bucket add contextverse https://github.com/orkcom-tech/scoop-bucket scoop install contextd

macOS / Linux — install script

curl -fsSL https://raw.githubusercontent.com/orkcom-tech/contextverse/main/scripts/install.sh | bash

Windows — install script

irm https://raw.githubusercontent.com/orkcom-tech/contextverse/main/scripts/install.ps1 | iex

Go

go install github.com/orkcom-tech/contextverse/cmd/contextd@latest

Also published as .deb / .rpm on Releases. The tap and bucket are ours — not submitted to homebrew-core or Scoop extras. Details in scripts/README.md.

Quickstart

The installer offers to run setup for you. To do it by hand:

contextd init

A guided wizard: pick a mode, and every choice is explained as you make it.

Mode	For	What it does
solo	One person, one machine	Local space at ~/.context. No server, no account.
client	Joining a team	Syncs a space from a server you have a token for.
server	Hosting for a team	Spaces, users, path ACL, audit. Opens a setup page.

Then, in any project:

contextd activate     # write entry points for every detected AI tool
contextd status       # what is wired, and where
contextd tui          # browse and edit the space full-screen

Change your mind later with contextd init --reconfigure.

How the context reaches your AI

contextd writes into each tool's own session-start slot. The slots are not equal, and we don't pretend they are:

Client	Slot	Nature
Claude Code	SessionStart hook	Live — re-reads the space every session
Cursor	.cursor/rules/*	Snapshot, refreshed on activate
Windsurf	.windsurfrules	Snapshot
GitHub Copilot	.github/copilot-instructions.md	Snapshot
opencode	AGENTS.md (marked block)	Snapshot; your own content is preserved
MCP clients	contextd mcp serve	Live tools over stdio
ChatGPT / web UIs	— none —	contextd export --format chatgpt, manual upload

Adding a new AI client is a PR to contextverse-templates, not a binary release.

Built for scripting

contextd is meant to be driven from CI, not just by hand.

contextd status --json | jq -r .mode
contextd file list --json
contextd space list --yaml        # server-side

Exit codes are meaningful, so a pipeline can tell "retry later" from "you called this wrong":

Code	Meaning
0	success
1	usage or local state error
2	network, auth or permission failure
3	compare-and-swap rejected — someone else wrote first

Shell completions: contextd completion bash|zsh|fish|powershell.

Three surfaces, one core

The CLI is primary. The terminal UI (contextd tui) and the server's web dashboard are presentations of the same operations — they never expose anything the CLI cannot do. Pick whichever is comfortable; the result and the permissions are identical.

Under the hood

• Context space model — layered identity / team / projects with entry files and freshness metadata.
• Derived link graph — the map comes from links your documents already contain, so backlinks, orphans and broken links are facts about what you wrote rather than inferences about it. space-index.md and team/space-map.md are generated from it, and paths pointing into a project's checkout are checked against the real files.
• Three surfaces over one listing — the CLI, the TUI and the web console answer "what is in this space" from the same code. They used to have a copy each, which agreed only on spaces the current build had just created.
• Pluggable storage backend — local filesystem, git, S3/MinIO, SQL. The backend is a dumb blob store with compare-and-swap; the core owns versioning, conflict and ACL semantics.
• Per-file versioning — Vault KV v2-style: integer versions shown as vN in CLI, TUI and web UI, with soft-delete, undelete and destroy.
• Pull-based sync — clients pull on activate; pushes are batched against the space head with CAS.
• Path-based access control — Vault-style policies and capabilities, deny-by-default.
• Stateless HA — no bespoke clustering: N stateless nodes behind a load balancer over an HA-capable backend.

Repository layout

cmd/contextd/     binary entrypoint
internal/         cli · tui · server · storage · syncclient · plugins · authz · audit · …
docs/             user documentation (MkDocs → GitHub Pages)
deploy/           systemd / launchd units, Docker and Helm templates
packaging/        scoop and winget manifest templates
scripts/          install.sh · install.ps1 · ci/

Related repo	Role
contextverse-templates	Context-space and client-integration templates (Apache-2.0)
homebrew-tap	Homebrew tap
scoop-bucket	Scoop bucket

Building from source

make build          # → bin/contextd
go test ./...
./bin/contextd init

Needs a recent Go toolchain. No cgo, no external services for the default build.

Documentation

Full user documentation: https://orkcom-tech.github.io/contextverse/

The source of truth is docs/ in this repository — edit it there.

pip install -r requirements-docs.txt
mkdocs serve

Contributing

Issues and design discussion are welcome. See docs/contributing.md. The project is early; a CLA/DCO lands when it formally accepts external code.

New AI client support does not need a code change — send a client-integration template to contextverse-templates.

Who builds this

ContextVerse is developed by ORKCOM. Copyright is held by Eduard Lugovtsov; ORKCOM is the project's home rather than a separate rights holder, and the licence names the person accordingly.

License

Business Source License 1.1 — source-available. Read it, self-host it, modify it, run it in production. You may not offer it to third parties as a hosted or embedded service that competes with ContextVerse. Each released version converts to Apache-2.0 four years after its release.

Templates are Apache-2.0. The managed ContextVerse Cloud is a separate, proprietary product.