Provenote MCP

Notebooklab

[Notebooklab hero showing grounded sources, auditable markdown, and reusable outputs in one workbench]

[Notebooklab quick-result overview showing the shortest repo-documented path from source import to auditable markdown download]

This illustrated overview is a repo-authored summary of the shortest documented path. It is intentionally not presented as a live product recording.

Canonical product path: messy long context -> structured insight -> note / research thread / draft -> inspectable outcome

That is the first door. MCP, starter bundles, distribution pages, and promotion assets are valuable second-ring surfaces, but they should not outrank the product path.

Agent-facing truth comes first: Notebooklab teaches an agent to read messy context, structure it, move it into note / research-thread / draft lanes, and only then carry that outcome workflow forward through the first-party notebooklab-mcp server. Public skills, host bundles, and registry packs are companion surfaces around that workbench, not the product root.

At A Glance

If you only want the shortest truthful filter before reading deeper, use this table first:

Start Here In 10 Seconds

If you only want the fastest honest map, use this:

Judge the workbench before you judge the side doors. MCP pages, starter bundles, distribution packs, and promotion assets matter, but they are second-layer surfaces around the main product path.

Why Notebooklab Exists

Most AI note tools make it easy to generate words and hard to verify where those words came from.

That gets even worse when the raw material is long and messy: a huge chat log, a copied forum thread, a meeting recap, or a web page pile you do not want to flatten into one more throwaway summary.

Notebooklab is built for the opposite direction:

• collect notes, documents, audio, and web content in one place
• structure long context before it disappears into another chat turn
• search and ask across that material without losing the source trail
• turn high-value source work into auditable markdown with integrity counters
• keep the whole workflow repository-documented and reproducible

Think of it like moving from a loose pile of research tabs to a workbench with labeled drawers, measuring tools, and a clean export lane.

If your problem starts as "I have too much messy context," the strongest current repo-backed answer is not another empty chat box. It is the path from long context to structured notes and then into reusable outcome objects.

Fastest Result Path

If you want one visible outcome before learning the whole workbench, keep the order this short:

1. import or open one source
2. open the source detail surface
3. run Auditable Markdown
4. download the inspectable result

That is the shortest repo-documented result path today. MCP, starter bundles, podcasts, and distribution surfaces stay second ring until this outcome path already makes sense.

If You Start With Messy Long Context

The strongest current first-entry path is:

Import the source-> Run Chat Knowledgeization-> Inspect the structured insight-> Continue it as a note or notebook research thread-> Keep draft work inside the notebook lane

In plain language: first turn the pile into labeled folders, then decide whether it belongs in your notes, your active research lane, or your notebook draft workflow.

What You Can Do In One Workbench

• Collect grounded context: bring in text, files, audio, and web sources instead of starting every session from a blank chat box.
• Structure long context before chatting more: use built-in transformations such as Chat Knowledgeization when the material starts as long chat logs, forum threads, meeting notes, or copied web discussions.
• Keep structured results moving: after long-context structuring, continue the result into a note, a notebook research thread, or the notebook draft lane instead of leaving the insight stranded.
• Search and think with structure: move between notebooks, ask/search flows, transformations, and model settings without hopping across disconnected tools.
• Create outputs with receipts: use the auditable markdown lane when you need stronger traceability than ordinary chat alone, or create notebook-level drafts plus export bundles when you need a reusable outcome object you can hand off.
• Hand outcomes off cleanly: download draft markdown when you only need the text, or export a richer draft bundle when you need metrics, claim/section review data, PID summaries, and source manifest context together.
• Go beyond text-only workflows: generate podcast-ready outputs and reusable transformations from the same source base.
• Bring it into coding agents through MCP: expose notebooks, sources, drafts, research threads, and auditable runs to Claude Code, OpenAI Codex, Cursor, and other MCP-capable hosts through a first-party MCP server.
• Run the same outcome lanes from a terminal/operator surface: use the first-party notebooklab CLI when you want notebook outcome inspection, auditable markdown, or research-thread-to-draft handoffs without treating MCP host setup as the only operator path.

If you only remember one first-entry rule, make it this:

messy long context-> structured insight-> note / seeded research lane / notebook research thread-> draft-adjacent notebook work

That is the product center. MCP host pages come after that path, not before it.

Ecosystem Boundary

Think of these surfaces like rooms around the main workshop:

• host guides are the side doors
• the CLI is the tool cart
• the main workbench is still the path from messy context to reusable outcomes

If you want the shortest public page that keeps those boundaries honest, start with docs/project-status.md.

If you want repo-owned install artifacts instead of prose-only setup pages, start with examples/hosts/README.md. That index points to the public-ready Claude Code, Codex, Cursor, and OpenCode starter bundles plus the OpenClaw-compatible bundles and the ClawHub submission pack.

If you want the standalone host-facing skill folders used for OpenHands/extensions or ClawHub-style submissions, start with public-skills/README.md.

If you want the full claim ladder instead of inferring it from scattered host pages, use docs/distribution.md.

Second-Layer Surfaces

Once the product path is clear, use one second-ring page instead of reading every host, registry, and distribution page in order.

The ordering is intentional: first the workbench, then the proof lane, then the carry-forward surfaces.

Product Truth Boundary

Ordinary chat and ask flows are a fast assistant surface for exploration and iteration.

When you need stronger traceability, the stricter product truth lane lives in auditable markdown and auditable-runs rather than ordinary chat alone.

Quick Result Path

This path optimizes for the first visible result, not full system mastery.

Think of it as the fastest repo-documented local proof loop, not a one-minute hosted trial.

In plain language: this is a 5 to 10 minute local proof loop, not a hosted one-minute trial.

1. Copy the local environment template and add the two fast-path values.

   bash

   cp .env.example .env

   Fast-path values:

   bash

   OPEN_NOTEBOOK_ENCRYPTION_KEY=change-me-to-a-secret-stringGEMINI_API_KEY=your-google-ai-studio-key

2. Start the default local stack.

   bash

   docker compose -f ops/compose/docker-compose.yml up -d --build

   On this Docker fast path, Compose injects the SurrealDB connection defaults and a known-good fast-path Gemini model for you. That is why the first visible result path only asks you to set the encryption key and Gemini key up front.

3. Open http://localhost:8502, let the app route you into the /sources workbench, then create or import a source and move into the source detail view.

4. Run Auditable Markdown to generate a downloadable markdown report with integrity counters.

5. Open a notebook and create a Draft if you want a reusable notebook-level outcome instead of a single-source artifact.

If you want the full walkthrough, go straight to docs/quickstart.md. If you want a zero-setup hosted product, this repository is not promising that experience today.

If you prefer a terminal-first operator path once the local stack is up, use the narrow first-party operator runbook in docs/runbooks/operator-cli.md. That surface is intentionally outcome-first and does not replace the workbench UI or the MCP compatibility pages.

If you want the ecosystem boundary before you read host pages, use docs/project-status.md.

Runtime Cleanup Operator Path

Notebooklab treats disk cleanup as a governed operator path, not an ad-hoc "delete big folders" exercise.

• Repo-local dependency root: apps/web/node_modules
  • This path is a repo-owned rebuildable dependency tree.
  • It is not a shared machine cache.
  • Clearing it is allowed only through the repo-local cleanup contract, and the rebuild command is cd apps/web && npm ci.
• Repo-related machine cache examples:
  • ${HOME}/.cache/notebooklab/...
  • ${HOME}/.cache/notebooklab/python/uv-cache
  • ${HOME}/.cache/notebooklab/playwright/ms-playwright
  • ${HOME}/.cache/notebooklab/ci-host/npm-cache
  • these are the only repo-specific machine-level download caches that should remain outside the checkout.
  • entrypoint wrappers now auto-run the repo-owned machine-cache cleanup lane so stale cache roots and oversized download caches do not silently accumulate forever.
• Repo-local CI/runtime residue:
  • .runtime-cache/venv/default
  • .runtime-cache/ci-host/...
  • these are repo-owned runtime or bootstrap surfaces and must stay inside the checkout rather than leaking into machine-wide cache roots.
• Shared tool caches remain advisory-only:
  • ${HOME}/.cache/uv
  • ${HOME}/.npm
  • ${HOME}/Library/Caches/ms-playwright
  • ${HOME}/Library/Containers/com.docker.docker
  • ${HOME}/.docker
  • these are related to the repo, but they are not repo-exclusive and must not be auto-cleared by Notebooklab entrypoints.
• Docker runtime surfaces:
  • local open-notebook-ci:* images
  • buildx builders / BuildKit containers
  • Docker build cache reported by docker system df -v
  • these belong to the Docker operator path, not the repo-local cleanup_runtime_cache.sh lane

Use the explicit operator flow:

make cleanup-operator-auditmake cleanup-operator-apply

That operator flow is intentionally split:

• cleanup_runtime_cache.sh handles repo-local runtime/cache surfaces
• cleanup_machine_cache.sh handles the remaining repo-related download caches under ~/.cache/notebooklab
• entrypoint wrappers (run_uv_managed.sh, run-playwright-managed.sh, and run_in_consistent_container.sh) invoke that machine-cache lane in apply mode so repo-specific external caches obey TTL/cap/root-cap policy by default
• docker-buildx-clean / docker-runtime-audit cover repo-related Docker builder and image surfaces

Real Result, Not Just Positioning

This is the fastest concrete proof path in the current repo surface:

The current UI and API evidence behind that lane is already public:

• the workbench exposes an Auditable Markdown panel with Run and Download Markdown actions
• the panel surfaces Coverage Rate, Missing, Duplicate, Uncited Claims, Unknown PID, and Unclassified
• the API publishes auditable-run routes behind the same workflow

Start from docs/proof.md if you want the file-and-route evidence behind those claims.

If you want a fixed, reproducible local proof loop, use the public proof pack.

Sanitized result shape

# Auditable Markdown Report
## Source- title: Example imported source- lane: auditable markdown
## Integrity Counters- coverage_rate: 0.98- missing_count: 0- duplicate_count: 0- uncited_claims_count: 0
## Output- downloadable markdown report- inspectable sections and claims- stronger traceability than a plain chat reply

This is a sanitized result shape, not a copied user artifact. The concrete product evidence for the lane lives in the current UI panel and auditable-run API surface.

Why People Star Notebooklab Now

• You want a repo-documented workbench, not another single chat pane. Notebooklab already combines sources, notebooks, search, transformations, podcasts, and an auditable output lane in one repo surface.
• You care about source-grounded writing that leaves receipts. The auditable markdown lane is a stronger hook than generic "AI notes" because it points people toward inspectable output, not only generated text.
• You want a repo worth revisiting, not a one-time curiosity. Releases, changelog, proof surfaces, and docs make it easier to track visible progress over time.
• You want a fork that acts like a real public product surface. Provenance still matters, but the current support, review, and stewardship boundary is repository-local and documented.

Why Watch Notebooklab Over Time

• The repo exposes proof surfaces, not only roadmap language. README, proof, changelog, release notes, and issue/discussion routing all work together as public evidence layers.
• The product shape is already broader than one feature spike. The current surface spans collection, search, research threads, notebooks, transformations, podcasts, source QA, and auditable outputs.
• The current fork boundary is explicit. That makes progress easier to interpret because support, review, and stewardship are not hidden behind upstream ambiguity.

Public Proof, Not Vibes

[Notebooklab public proof stack showing product proof, workflow proof, trust proof, and growth proof]

If you want proof before commitment, start here:

Release visibility is one public signal, not automatic proof that the latest release-event build and asset set are clean.

Docs By Goal

Use It With Coding Agents

Notebooklab ships a first-party MCP server so you can bring the same outcome objects you use in the web workbench into coding-agent hosts.

If your starting problem is messy long context, start with Long Context first and come back here when you want to carry those outcome objects into a host.

Current public fit:

• use Notebooklab with Claude Code through MCP
• use Notebooklab with OpenAI Codex through MCP
• use Notebooklab with Cursor through MCP
• use Notebooklab with OpenCode through MCP

Current boundary:

• this is a compatibility surface, not an official vendor partnership
• vendor names are used descriptively to show host targets
• the product center is still Notebooklab's auditable/source-grounded outcome path, not the MCP control plane by itself

Start with docs/mcp.md if you want the MCP overview, then choose the host-specific page that matches your agent runtime.

If you want checked-in starter artifacts instead of only setup pages, go straight to examples/hosts/README.md. That index links the public-ready Claude Code, Codex, Cursor, and OpenCode starter bundles plus the OpenClaw-compatible bundles and the ClawHub submission pack.

Product Surface At A Glance

The main workbench currently spans:

• Collect: sources and notebooks
• Process: ask/search and transformations
• Create: podcasts, notebook drafts, and auditable markdown
• Manage: models, settings, advanced controls, and repo-local governance

[Notebooklab architecture view showing the UI, API, core runtime, and verification surfaces]

Best Fit And Not The Best Fit

Notebooklab is a strong fit if you want to:

• turn notes, documents, audio, and web sources into inspectable markdown outputs
• run a repository-documented research workbench instead of depending on a hosted SaaS
• evaluate a NotebookLM-like surface that already includes notebooks, search, podcasts, and auditable output paths
• follow a repo that publishes proof surfaces, release surfaces, and stewardship boundaries in public

Notebooklab is probably not the best fit if you want:

• a zero-setup hosted product with no local runtime or Docker involvement
• a multi-provider runtime by default today
• a tiny single-purpose summarizer or export-only utility
• a repo that hides fork lineage or blurs who currently owns support and review

Brand And Lineage

Notebooklab is the public identity of this repository.

This repository is a deep, productized fork of the upstream Open Notebook project.

Upstream lineage still matters for provenance, but the current support, review, release, and collaboration surface is repository-local to this checkout.

That means two things can both be true at once:

• Notebooklab is the brand people should remember
• upstream lineage still belongs in the evidence layer, not the headline

The repo-local stewardship and trust boundary is anchored in NOTICE.md, MAINTAINERS.md, SUPPORT.md, SECURITY.md, and CONTRIBUTING.md.

Local Cleanup Operator Path

When disk pressure comes from local Docker builders plus repo-local rebuildables, use one explicit operator path instead of guessing which command owns which surface:

make cleanup-operator-dry-runmake cleanup-operator-rebuildablemake cleanup-operator-aggressive

That split is intentional:

• docker-buildx-clean handles local Buildx/builder residue
• tooling/scripts/ops/cleanup_runtime_cache.sh handles repo-local runtime and rebuildable surfaces
• apps/web/node_modules remains a repo-local rebuildable dependency root, not a machine cache
• repo-related machine caches still stay review-first instead of being auto-cleared by repo automation, and they are limited to download-cache surfaces only

Community

• Questions and product discussion: GitHub Discussions
• Reproducible bugs and scoped improvements: GitHub Issues
• Contribution flow: CONTRIBUTING.md
• Support boundary: SUPPORT.md
• Security reporting: SECURITY.md
• Collaboration conduct: CODE_OF_CONDUCT.md
• Current review routing: CODEOWNERS
• Stewardship and provenance: NOTICE.md and MAINTAINERS.md

License

This fork continues to distribute upstream-licensed material under LICENSE.