Knowledge Hub MCP server
Commit Nov 13, 2026Lets an AI client ingest, manage, and review canonical knowledge stored in a local-first SQLite database on macOS.
Also known as NewBww/knowledge-hub
- Stars
- 0
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 13, 2026
- License
- MIT
What is Knowledge Hub MCP?
Knowledge Hub lets an AI client ingest, manage, and review canonical knowledge stored in a local-first SQLite database on macOS.
Summary generated from the project's README
About Knowledge Hub MCP server
Managing and reviewing canonical knowledge within a local-first SQLite database on macOS is what Knowledge Hub enables. It functions as a reading room where users and agents collaborate to maintain a shared repository of information. The system uses a Library for reading subjects and a Warehouse for inspecting evidence, ensuring that all changes undergo a review process before being accepted.
Developers can use this via a client-neutral stdio MCP connection. While reading and searching do not require a model account or inference credentials, an optional Curator or a personal MCP agent can be used to prepare changes. The setup requires Python 3.12 and uv on macOS to run from the source. It is designed for users who need explicit human approval for knowledge updates, as neither the Curator nor an agent can approve its own work.
What you can do with it
- 1Verify knowledge changesAsk the AI to present pending knowledge proposals and evidence for human review before they are committed to the database.
- 2Search local informationQuery the local SQLite database to find specific facts or subjects without needing an active model account.
- 3Ingest scoped local filesProvide the agent with specific files or folders using a picker to create durable Curator jobs.
- 4Inspect source evidenceRequest the AI to look into the Warehouse to find the specific evidence supporting a piece of knowledge.
Written from the project's README. Every use case is checked against a quote from it.
README
View on GitHubKnowledge Hub
Unfinished source-only DRAFT — 0.3.0 unreleased. This is not a finished product or a cleared binary release.
A local-first macOS reading room for knowledge that people and agents maintain together.
Read in Library. Inspect evidence in Warehouse. Review before accepting changes.
Knowledge Hub stores canonical knowledge in SQLite on your Mac. Reading and search do not require a model account. An optional bounded Curator or your own MCP agent can prepare changes; neither can approve its own work.
Start here
- Install and capture your first knowledge
- Library, Warehouse and human review
- Privacy, permissions and retention
- Connections and their actual limits
- Use your own agent
- Troubleshooting and keyboard access
- Backup and migration
An explicitly unfinished source-only DRAFT may be shared without completing binary acceptance or frozen-payload licensing. Before any public push, review the exact allowlisted bytes for privacy, secrets, authorship/origin and intentional attribution; DRAFT visibility is still public. This documentation does not resolve ownership or legal clearance. Protected-main merge and finished release still require actual green CI, applicable acceptance checks and relevant license gates; no force push or protection bypass is authorized.
macOS installation
Only install a candidate whose release notes identify its version, architecture, digest and signing limitations. The repository's presence is not evidence that a distributable binary has cleared release gates. Follow installation and verification. Current packaging uses an ad-hoc local-test signature, not Apple Developer ID signing or notarization. Binary dependency review, native updater acceptance and publication are separate gates; see releasing.
The default data directory is ~/Library/Application Support/Knowledge Hub, outside the app. Quit before replacing the app and back up your data first. Removing the app does not delete knowledge or account state.
Source quickstart
On macOS, install Python 3.12 and uv, then run from the checkout:
uv sync --extra desktop --group dev --locked
uv run --locked --extra desktop --group dev knowledge-hub
For an isolated trial, append --data-dir "$HOME/Knowledge-Hub-Trial"; use a new directory, not an existing personal knowledge store. No hosted database, Hermes account or inference credential is required to launch and read. See Contributing for the full test setup, including Chrome.
What works, and what is not promised
- One shared Warehouse; no required projects. Library books are readable subjects, not containers that own facts. Knowledge can exist without books.
- Local capture of text and bounded attachments; Activity distinguishes saved work from actual processing. PDF/DOCX text extraction is bounded; images are validated, not OCR-read.
- Pending knowledge and reading-surface proposals, revision checks, evidence and explicit human approval/rejection. Dependency impact covers registered relationships, not every possible contradiction.
- Client-neutral stdio MCP with progressive retrieval and explicit omissions. Remote clients can forward anything they read: local storage is not a provider privacy guarantee.
- Settings separates connection choice, model selection, sign-in, synthetic checking, disclosure consent, capture access and processing enablement. ChatGPT/Codex execution remains blocked; other routes have version/capability limits.
- A native mixed file/folder picker registers scoped, read-only local context. Select sources explicitly for each durable Curator job; this is not bulk folder upload. Jobs stop on Quit and recover eligible work on reopen after revalidation; no closed-app processing is configured; canonical changes still require human review. See the workflow.
- Arbitrary source polling, automatic connector imports, universal freshness and completed cross-book follow-up/outbox processing are not established capabilities. The strict-CSP source bridge repair is included. Later frozen QA opened the real native panel and verified cancellation, but positive mixed selection/grant/revoke acceptance remains blocked. Public binary gates remain open; see current evidence.
For the implementation/roadmap boundary see Product, the roadmap and release backlog, and documentation status. Source metadata is authoritative for the version; an untagged checkout is not a release.
Developer and maintainer reference
Developer quickstart · Code map · Architecture and API · Design · Contributing · Security · Changelog · Release maintenance · Wiki maintenance
Application source is MIT licensed. Dependency licenses and frozen-bundle obligations are separate; third-party notices do not by themselves certify a binary for redistribution. User knowledge is not licensed by this repository.
Knowledge Hub MCP questions
Does reading and searching require a model account?
No, reading and searching do not require a model account.
Where is the data stored on macOS?
The default data directory is located at ~/Library/Application Support/Knowledge Hub.
Can an AI agent approve its own changes to the knowledge base?
No, while an MCP agent can prepare changes, neither the Curator nor an agent can approve its own work.