OpenClaw Notebook Tools

Cell-aware Jupyter notebook editing for OpenClaw agents, without raw JSON surgery.

OpenClaw plugin: read and edit Jupyter notebooks (.ipynb) with structured, cell-aware tools instead of raw nbformat JSON.

After install, the agent gets all 9 tools — no extra configuration needed. Mutating tools are guarded by atomic writes and stale-edit checks rather than by a per-tool opt-in.

Tool	Side effect	Purpose
notebook_read	No	Full or targeted structured read
notebook_list_cells	No	Cheap index — id, type, preview, counts
notebook_search	No	Find cells by source text / regex
notebook_validate	No	Validate notebook structure and report facts
notebook_create	Yes	Create a new notebook from structured cells
notebook_edit_cell	Yes	Replace a cell's source / type
notebook_insert_cell	Yes	Insert a new cell at position
notebook_delete_cell	Yes	Delete a cell by id or index
notebook_clear_outputs	Yes	Clear outputs (all or single cell)

Why?

OpenClaw's built-in read / write / edit tools treat .ipynb files as opaque JSON. The agent has to manipulate raw nbformat — error-prone, token-hungry, and easy to corrupt. This plugin provides a cell-aware API with:

• Atomic writes (temp file + fsync + rename) so a crash mid-edit can never corrupt the user's notebook.
• Stable cell ids (nbformat 4.5+, backfilled in memory for older notebooks) so multi-step edits don't drift after inserts/deletes.
• Stale-edit guards (expected_source_sha256, expected_file_mtime_ms) so concurrent edits in JupyterLab/VS Code can't be silently overwritten.
• Output-aware truncation so reading a notebook full of large outputs doesn't blow your token budget.

Safety model

This plugin intentionally exposes all nine tools as always-on once installed. That keeps notebook work ergonomic for agents: they can inspect, edit, clean, and validate notebooks without a second allowlist step.

The trust boundary is implemented in the tool behavior instead:

• Mutating tools use sequential execution so concurrent tool calls do not race.
• Writes are atomic: save to a sibling temp file, fsync, then rename.
• Mutations validate notebook structure before replacing the original file.
• Stale guards (expected_source_sha256, expected_file_mtime_ms) prevent silent overwrites when a notebook changed after the agent inspected it.
• Editing a code cell clears its outputs and execution count, avoiding stale result displays.
• The plugin never executes notebook code. It only reads and edits files.
• Agents should run notebook_validate after mutations before reporting success.

Install

Install from ClawHub:

openclaw plugins install clawhub:@lightbridge-ks/openclaw-notebook-tools
openclaw plugins list
openclaw gateway restart

For local development, install from a packed tarball (not from the source directory — see "Why pack first?" below):

pnpm install
pnpm pack:plugin                   # builds, refreshes plugin metadata, then runs `npm pack`
                                   # → lightbridge-ks-openclaw-notebook-tools-0.1.0-beta.2.tgz

openclaw plugins install ./lightbridge-ks-openclaw-notebook-tools-0.1.0-beta.2.tgz
openclaw plugins list              # confirm `notebook-tools` is listed
openclaw gateway restart

Why pack first?

openclaw plugins install <dir> recursively copies the source directory into ~/.openclaw/extensions/<plugin>/ and then runs npm install there. With a pnpm-managed dev tree, that copies hundreds of MB of symlinked node_modules and breaks the install.

npm pack respects the files whitelist in package.json, producing a small tarball with runtime JS, the plugin manifest, public docs, and the shipped skill. OpenClaw extracts the tarball into a clean staging directory and runs npm install against just our runtime deps (typebox, nanoid).

Develop

pnpm install
pnpm test         # vitest watch mode
pnpm test:run   # single run (CI mode)
pnpm typecheck    # tsc --noEmit
pnpm build        # tsc → dist/
pnpm plugin:build # build + regenerate OpenClaw plugin metadata
pnpm plugin:check # build + fail if generated metadata is stale
pnpm plugin:validate
npm pack --dry-run --json
clawhub package publish "$(pwd)" --owner Lightbridge-KS --version 0.1.0-beta.2 --dry-run --json

The codebase follows Clean Architecture:

• src/nb/ — pure notebook domain (no SDK imports), unit-testable in isolation.
• src/tools/ — adapter layer: TypeBox schemas, error mapping, tool registration.
• index.ts — composition only, wires tools into defineToolPlugin.