openhoard

"Open the Acme Q3 deck." · "What CSVs was I working on yesterday?" · "Find the onboarding checklist HR wrote last spring." · "Who can see payroll, and why?"

OpenHoard replaces the folder tree with an AI-managed, governed file layer. Every file is indexed, tagged, summarized and access-controlled, so people and their AI agents find things by meaning, not by path. Every action is checked by policy and written to a tamper-evident audit log.

Status: pre-alpha, designed in the open. What the headlines say is in main today, tested against a simulated SharePoint tenant; the first real site is next. What is planned is listed as such, below them. The roadmap has the order.

The headlines

Planned (on the roadmap, not in main):

• Governed sharing: "Share the forecast with Acme, read-only, until Oct 31", with external links that expire by default.

• Cloud buckets and Git: index and move shared work in S3 or Azure Blob; link GitHub repositories. (The files OpenHoard keeps itself, uploads and mail-in, can already live in your own S3 or Azure Blob bucket, set up from the admin Storage page.)

• SFTP is over: send customers a link, and they drag in 1 GB+ files from any browser, with no account and no client.

• No more Final_v2_FINAL.docx: AI that names files, spots the real final version and warns you before you send an old one.

• 30-day undo on everything, and HIPAA / SOC 2 / legal-hold packs.

• Plugins in a sandbox, with the capabilities they declare and nothing more.

• Checked live with Claude, ChatGPT and Microsoft Copilot. OpenHoard speaks standard MCP with OAuth 2.1; it is tested against the protocol, not yet against those apps.

Related MCP server: Glean Remote MCP Server

Why your company needs this (the slide for your boss)

The folder tree was designed for filing cabinets. It's failing modern teams:

• 54% of US office workers say they waste time searching for files in disorganized systems (Elastic / Wakefield Research via TechRepublic).

• 77% of organizations had at least one Microsoft 365 governance incident in the past year (ShareGate survey via WindowsForum).

• 38% had former employees or guests keep access they should have lost, and 26% had sensitive content reach the wrong people (same survey).

• AI assistants now read everything a user can access, so oversharing that used to be hidden is now one prompt away.

OpenHoard's pitch in one line: keep your storage, lose the chaos. Findable files, provable access control, and AI you can actually trust with company data.

• Deploys like any managed app: SSO (Entra ID, Google, or any OIDC provider) and SCIM today; silent sign-in and MDM packages planned.

• Least privilege by design: SharePoint access via Sites.Selected; AI clients on an allowlist; plugin sandboxing planned.

• One policy layer for everything OpenHoard indexes, readable as "who can see this and why".

• Security-first AI: prompt injection is in the threat model from day one. Text in a file can never authorize an action, and files that look like they try are kept from AI assistants.

• No lock-in: open source, self-hostable; the audit history exports today, files and tags export planned.

The IT overview deck (PDF) is a September 2026 snapshot, kept for its diagrams of the architecture, data flow, tags and security. Where it differs from this README (what OpenHoard is for, which sources it indexes, status, roadmap), this README is current.

How it works

            ┌──────────────── Trusted core (small, audited) ────────────────┐
            │ identity · permissions/policy · index/search · summaries ·    │
            │ audit · plugin sandbox                                        │
            └───────────────────────────────▲───────────────────────────────┘
                                            │ versioned contracts + capability manifests
   connectors · enrichers · policy/vocabulary packs · skills · client integrations

The core makes every trust decision. Plugins extend what OpenHoard can reach and understand; they can never decide who gets access. Deep dives: architecture · threat model · roadmap.

Repository layout

Try it

On your own machine, over your own folder, in a few commands. Node 24, pnpm and git are all the server needs; no Docker, no database to install:

git clone https://github.com/openhoard/openhoard
cd openhoard
pnpm install
pnpm build
node apps/cli/dist/bin.js init --solo

init --solo sets up one person, a folder (OpenHoard in your home folder, made if missing; or --folder <an existing folder>, the path written out) and a starter policy pack. It then prints the two commands that are left: one connects Claude Desktop (connect claude-desktop), one starts the server. Follow what they print (a sign-in link to open, Claude Desktop to restart, Allow to press), put a file in the folder, and ask Claude for it.

• Search works with no AI key. For summaries, set OPENHOARD_MODEL_CLAUDE_API_KEY in the environment (never in a config file).

• No Claude Desktop (Linux has none), or you want claude.ai or your phone: start it with tunnel instead (needs cloudflared) and follow what it prints. Add files in a browser at <its address>/app/.

The details are in apps/server/README.md ("One person on one machine"). This is pre-alpha: it never writes to the folder you name, but expect rough edges.

For a team on Microsoft 365, init --team sets up a server that signs people in through Entra and indexes a SharePoint site in place. docs/pilot.md goes from nothing to three people finding the site's files through Claude; docs/admin.md is the admin's guide after that, and docs/clients.md connects Claude and other AI clients.

The CLI validates plugin manifests against the published schema, the first contract every plugin must meet.

node apps/cli/dist/bin.js manifest validate connectors/example/openhoard.plugin.json

Join the hoard

We're building this in the open and want the community to own the edges. Start with CONTRIBUTING.md. Connectors, enrichers, packs and skills are the best places to begin. Found a security issue? See SECURITY.md.

License

Apache-2.0 · "OpenHoard" and the dragon-eye logo are trademarks of the OpenHoard project.