Notebook Intelligence (NBI) is an AI coding assistant and extensible AI framework for JupyterLab. It adds chat, inline edit, auto-complete, and an agent that can drive notebooks — backed by GitHub Copilot, an OpenAI-compatible or LiteLLM-compatible endpoint, local Ollama models, or Anthropic's Claude Code CLI.
NBI is free and open-source. Connect it to a free or paid LLM provider of your choice — GitHub Copilot, any OpenAI- or LiteLLM-compatible endpoint, Ollama (local), or Anthropic Claude (via the Claude Code CLI). Provider charges, when applicable, are paid directly to the provider.
- Requirements
- Quick start
- Concepts
- Feature highlights
- Configuration
- Built-in tools
- Model Context Protocol (MCP) support
- Rulesets
- Claude Skills
- Claude MCP Servers
- Claude Plugins
- Chat feedback
- Configuration readiness
- Performance diagnostics
- Documentation
- Further reading
- Roadmap
- License
- Python 3.10+
- JupyterLab 4.x
- Node.js — only required for Claude mode (the Claude Code CLI) and for MCP servers that launch via
npx. - A fresh virtualenv or conda env is recommended so NBI doesn't conflict with system Python.
pip install notebook-intelligence
jupyter lab # restart JupyterLab if it was already runningAfter restart:
- Click the NBI icon in the left sidebar to open the chat panel.
- Open NBI Settings (gear icon in the chat panel, or Settings → Notebook Intelligence Settings).
- Sign into your provider — for GitHub Copilot, click Sign in; for an OpenAI- or LiteLLM-compatible endpoint, paste an API key; for Ollama, point at your local daemon. To use Claude, enable Claude mode (see below).
- Type a message in the chat panel and press Enter.
If the panel stays empty or login does nothing, see Troubleshooting.
A short glossary you'll see referenced throughout these docs.
- LLM Provider — the service that runs the model. NBI ships with four provider adapters: GitHub Copilot, OpenAI-compatible, LiteLLM-compatible, and Ollama. Anthropic Claude is available through Claude mode, not as a top-level provider.
- Chat Participant — a
@mention-able persona inside the chat panel (@workspace,@mcp, …). Participants route the request to a specific tool surface. - Default mode vs Claude mode — Default uses the configured LLM Provider for chat, inline chat, and auto-complete. Claude mode uses the Claude Code CLI for the chat panel (gaining its tools, skills, MCP servers, and custom commands) and Claude models via the Anthropic API for inline chat and auto-complete. Requires the Claude Code CLI on
PATH. - Claude Code vs the Anthropic API — the Anthropic API (
api.anthropic.com) is the HTTPS endpoint NBI calls directly for inline chat and auto-complete in Claude mode. Claude Code is Anthropic's local CLI agent that NBI shells out to for the chat panel; it talks to Anthropic itself. - MCP — Model Context Protocol. A way for the LLM to call out to external tools (read files, hit APIs, run scripts).
- Ruleset — markdown files in
~/.jupyter/nbi/rules/that get injected into the system prompt to enforce conventions, coding standards, or domain rules. - Skill: a directory under
~/.claude/skills/(or<project>/.claude/skills/) holding aSKILL.mdplus helper files. Claude can invoke it like a callable plugin scoped to a workspace. - Claude plugin: a unit packaged for
claude plugin install, distributed through a marketplace (typically a GitHub repo that publishes a manifest of plugins). Distinct from NBI's own labextension; plugins run inside Claude Code sessions.
NBI provides a dedicated mode for Claude Code integration. In Claude mode, NBI uses the Claude Code CLI for the chat panel, and Claude models (via the Anthropic API) for inline chat and auto-complete suggestions. This brings Claude Code's tools, skills, MCP servers, and custom commands into JupyterLab.
Configure via the NBI Settings dialog (gear icon in the chat panel, or Settings → Notebook Intelligence Settings). Toggle Enable Claude mode, then:
- Chat model — the Claude model used for the chat panel and inline chat.
- Auto-complete model — the Claude model used for auto-complete suggestions.
- Chat Agent setting sources — user, project, or both, mirroring Claude Code's settings.
- Chat Agent tools — which tool sets to activate. Claude Code tools are always on. Jupyter UI tools are NBI's own (authoring notebooks, running cells, etc.).
- API key and Base URL — point at Anthropic or a self-hosted endpoint.
- Show turn usage: off by default. When on, each completed turn ends with a one-line italic footer giving the turn's duration and token counts. A
$cost figure appears only when NBI is running against a direct Anthropic API key on Anthropic's own endpoint: the SDK prices from the CLI's built-in public list rates, which are not the marginal cost on a subscription login and not the rate on a custombase_url, so NBI omits a number it cannot stand behind rather than showing a misleading one.
Model defaults. With no model selected, the chat model resolves to the newest Sonnet (claude-sonnet-5 today) and auto-complete to the newest Haiku, both matched against the model list fetched from the configured endpoint rather than a hardcoded ID. If a configured ID no longer exists, NBI falls back within the same tier by version rather than alphabetically, so claude-sonnet-99 resolves to the newest Sonnet. The fetched list is cached per endpoint, so changing Base URL does not reuse the previous endpoint's models. Auto-complete responses are capped at 1024 tokens, since a suggestion is at most a few dozen lines; NBI_CLAUDE_INLINE_COMPLETION_MAX_TOKENS overrides the cap and is clamped to [1, 4096].
If the Claude Code CLI is on PATH, NBI launches it automatically. To override the location, set the NBI_CLAUDE_CLI_PATH environment variable before starting JupyterLab.
In Claude mode the chat input footer shows a shield-icon button (to the left of the send button) that sets the agent's permission mode for the chat panel, matching the modes in Claude Code and the Claude VS Code extension. Click it to choose:
- Default: every tool call the agent wants to run goes through NBI's confirmation prompt. You approve or reject each one. This is the starting mode.
- Accept Edits: file edits the agent makes apply without a per-edit prompt; other tool calls (running commands, etc.) still go through the confirmation prompt. Useful for iterative work where you trust the edits but still want a gate on everything else.
- Plan: the agent researches and proposes a plan without making any changes, then presents it for approval. Approving runs the plan and returns the selector to Default; rejecting keeps it planning. This replaces the old
/enter-plan-modeslash command. - Bypass Permissions: NBI's confirmation prompt is skipped for every tool call, including the Claude Code CLI's own Bash / Write / Edit running in the agent subprocess. The agent runs everything with your full account access and no confirmation, and any untrusted content it reads can steer what it runs. See the gating notes below.
Default, Accept Edits, and Plan switch the moment you pick them. The selected mode travels with each message you send and is applied to the agent before the turn runs; switching mid-conversation takes effect on your next message.
Bypass Permissions never persists: starting a New chat session (or /clear) always drops it and it has to be re-armed manually. The other modes carry over across a reset, and a fresh Claude client's starting mode is Default (or an administrator's managed permissions.defaultMode).
Choosing Bypass Permissions does not arm it immediately. It opens a confirmation step; only after you confirm does bypass take effect, and while it is active the shield turns into a red warning icon as a persistent indicator. Bypass must be re-armed each session: starting a new chat or restarting the Claude client drops back to Default. And because the server re-checks the requested mode on every message, an armed bypass can never outlive a policy that an administrator has since turned off.
Admin gating. Bypass Permissions is off by default and hidden from the selector unless an administrator enables it: it is governed by the claude_bypass_permissions policy (NBI_CLAUDE_BYPASS_PERMISSIONS_POLICY), one of three admin policies that default to force-off rather than user-choice (the others gate the experimental ACP agent mode and ACP full access). The requested mode is also clamped on the server for every message, so the gate can't be bypassed by a hand-crafted request. Independently, NBI honors Claude Code's enterprise managed settings: permissions.disableBypassPermissionsMode removes the option regardless of the NBI policy, and permissions.defaultMode sets the selector's starting mode (Bypass excepted, since it never auto-arms). See Allowing Bypass Permissions in the admin guide.
The /enter-plan-mode and /exit-plan-mode slash commands still work if typed but are no longer offered in autocomplete; the selector replaces them and will retire the commands in a future release.
When Claude mode is on, the chat sidebar shows a history icon next to the gear. Click it to list the Claude Code sessions recorded for the current working directory (the same transcripts the Claude Code CLI stores under ~/.claude/projects/). Selecting a session reconnects via resume, so the next message you send continues that transcript with full prior context. A New chat session button next to the gear restarts the SDK client without typing /clear.
Long Claude turns surface an elapsed-time counter and a heartbeat-driven pulse with a "may be slow" copy flip after 30 seconds. Each tool the agent runs shows up as a persistent status card with a kind icon and a live in-progress / done / failed state; edits carry an inline diff, and a run of consecutive calls collapses into one expandable group, so the sidebar reflects what the agent is doing rather than appearing stuck.
In Claude mode, workspace files attached as chat context arrive as @-mention pointers rather than inlined file contents. Claude's Read tool fetches them on demand, which means images, large files, and notebooks (cell-aware) now work where the older content-injection path silently truncated or skipped them.
Enable Show usage after each turn in the Settings dialog to append a small footer under each Claude-mode reply with the turn's duration, token counts (with a cached-input breakdown), and cost — for example 12.3s · 45.2K in (38.1K cached) / 1.2K out · $0.0842. It is off by default.
The cost figure comes from the Claude Code CLI, priced from Anthropic's public list rates. It is shown only when NBI is configured with a direct Anthropic API key on the default endpoint, and is omitted otherwise — on a subscription login (where the marginal cost is effectively $0), an enterprise-negotiated contract, or a custom Base URL (proxy, gateway, or non-Anthropic endpoint) — since the list-price figure would not match your actual billing. Duration and token counts are always shown.
When the Claude CLI is on PATH, the JupyterLab launcher (the panel that opens with new tabs) shows a Claude Code tile alongside the standard kernel launchers. Clicking it opens a session picker; search across past transcripts and resume one in a fresh terminal, or start a new session in the file browser's active subdirectory. Session IDs are copyable from the picker for paste into a claude --resume <id> command.
When any of the following CLIs are on PATH, the launcher adds a tile for each. Clicking a tile opens a terminal at the file browser's current directory and runs the CLI:
- opencode (override path with
NBI_OPENCODE_CLI_PATH) - Pi (override path with
NBI_PI_CLI_PATH) - GitHub Copilot CLI (override path with
NBI_GITHUB_COPILOT_CLI_PATH) - OpenAI Codex (override path with
NBI_CODEX_CLI_PATH)
Tiles add and remove themselves as CLIs become available or unavailable; they do not require Claude mode. Clicking any tile (or New Session from the Claude resume dialog) prompts for a start directory, so the terminal opens where you want rather than always at the file-browser cwd.
In Agent mode, the built-in AI agent creates, edits, and executes notebooks for you interactively. It can detect issues in cells and fix them.
Use the sparkle icon on the cell toolbar or the keyboard shortcut to show the inline chat popover.
Ctrl+G / Cmd+G opens the popover. Ctrl+Enter / Cmd+Enter accepts the suggestion. Esc closes it. The accept shortcut overrides JupyterLab's default run cell binding only while the popover is open — outside the popover, Ctrl+Enter / Cmd+Enter still runs the active cell.
Auto-complete suggestions are shown as you type. Tab accepts. NBI provides auto-complete in code cells and Python file editors.
You can paste or attach images alongside a chat prompt — the image goes to the model as input when the active model supports vision.
Right-click a cell output (or hover for the toolbar) to send it straight into the chat as context:
- Explain cell errors — surfaces a "Troubleshoot errors in output" entry on cells that raised; opens a chat turn with the traceback attached.
- Ask about cell outputs — attaches the output as structured context for a follow-up question. Includes images for vision-capable models.
- Show output toolbar — the floating toolbar above each output with quick Explain / Ask / Troubleshoot actions.
Each is per-user toggleable from Settings (saved as enable_explain_error, enable_output_followup, enable_output_toolbar in config.json, default on) and admin-lockable via NBI_EXPLAIN_ERROR_POLICY / NBI_OUTPUT_FOLLOWUP_POLICY / NBI_OUTPUT_TOOLBAR_POLICY.
Active notebooks show a sparkle icon on the toolbar. Click it to open a popover that scopes the generation request to that specific notebook — handy for multi-notebook sessions where you don't want the chat sidebar to compete for context.
Notebook creation, cell insertion, and inline generation follow the notebook you are actually in rather than assuming Python. NBI resolves the active notebook's kernel and language and carries both through the chat context, the built-in tools, and ruleset scoping, so generated code matches the target notebook, and a rule scoped to languages: ['r'] or kernel_names: ['ir'] applies where you would expect (including in inline generation, which previously did not receive the kernel).
When the agent needs a notebook in a language the current context has not established, it calls a list-available-notebook-kernels tool and picks from the kernelspecs that are actually installed instead of guessing a name. Asking for a kernel that is not installed now fails with a clear error rather than silently producing a Python notebook.
NBI reloads open document tabs when their files change on disk, so edits an AI agent makes via its Read/Write tools appear in the editor without a manual refresh. Tabs with unsaved local edits are skipped so user work is never clobbered. Toggle via the NBI Settings dialog → External changes → "Refresh open files when changed on disk" (default on).
Configure your provider, model, and API key from NBI Settings — the gear icon in the chat panel, the /settings chat command, or the JupyterLab command palette. For background, see the provider blog post.
NBI saves configuration at ~/.jupyter/nbi/config.json. It also supports an environment-wide base configuration at <env-prefix>/share/jupyter/nbi/config.json — organizations can ship default configuration there, and user changes save as overrides on top.
These config files store provider, model, and MCP configuration. API keys for custom LLM providers are also stored here in plaintext — never commit ~/.jupyter/nbi/config.json to git, share it, or sync it across users. If a key leaks, rotate it at the provider immediately.
Manual edits to
config.jsonrequire a JupyterLab restart to take effect. Edits via the Settings dialog are picked up live.
Most settings panel toggles can be locked by org administrators. Two shapes:
Boolean policies use the *_POLICY suffix and accept three values: user-choice (default — user toggles freely), force-on (locked enabled), force-off (locked disabled). When forced, the panel control is disabled with a "Locked by your administrator" tooltip and any client-side write is ignored.
| Env var | Locks the Settings panel control for |
|---|---|
NBI_EXPLAIN_ERROR_POLICY | "Explain cell errors" |
NBI_OUTPUT_FOLLOWUP_POLICY | "Ask about cell outputs" |
NBI_OUTPUT_TOOLBAR_POLICY | "Show output toolbar" |
NBI_CLAUDE_MODE_POLICY | "Enable Claude mode" |
NBI_ACP_MODE_POLICY | "Enable ACP mode" (experimental agent mode over the Agent Client Protocol, with Codex as the first agent type; defaults to force-off, set user-choice to allow it) |
NBI_ACP_FULL_ACCESS_POLICY | "Full access" for the ACP agent: run tools without asking (defaults to force-off, so the agent asks before anything beyond trusted read-only commands; user-choice lets users opt into unattended runs) |
NBI_CLAUDE_CONTINUE_CONVERSATION_POLICY | "Remember conversation history" |
NBI_CLAUDE_CODE_TOOLS_POLICY | "Claude Code tools" |
NBI_CLAUDE_JUPYTER_UI_TOOLS_POLICY | "Jupyter UI tools" |
NBI_CLAUDE_SETTING_SOURCE_USER_POLICY | Setting source: User |
NBI_CLAUDE_SETTING_SOURCE_PROJECT_POLICY | Setting source: Project |
NBI_STORE_GITHUB_ACCESS_TOKEN_POLICY | "Remember my GitHub Copilot access token" |
NBI_SKILLS_MANAGEMENT_POLICY | The Skills tab (force-off hides it and 403s the API; also disables the managed-skills reconciler) |
NBI_CLAUDE_MCP_MANAGEMENT_POLICY | The Claude-mode MCP Servers tab (force-off hides it and 403s /claude-mcp/*; independent of the non-Claude MCP Servers tab) |
NBI_CLAUDE_PLUGINS_MANAGEMENT_POLICY | The Claude-mode Plugins tab (force-off hides it and 403s /plugins/*) |
NBI_CLAUDE_BYPASS_PERMISSIONS_POLICY | "Bypass Permissions" in the Claude permission-mode selector (defaults to force-off, as do NBI_ACP_MODE_POLICY and NBI_ACP_FULL_ACCESS_POLICY; user-choice exposes the option, which the user still arms per session) |
NBI_TERMINAL_DRAG_DROP_POLICY | Terminal drag-drop file attach feature |
NBI_REFRESH_OPEN_FILES_ON_DISK_CHANGE_POLICY | "Refresh open files when changed on disk" |
The first three also have matching traitlets on NotebookIntelligence (explain_error_policy, output_followup_policy, output_toolbar_policy); add the others as needed in the same shape:
c.NotebookIntelligence.claude_mode_policy="force-on"c.NotebookIntelligence.claude_jupyter_ui_tools_policy="force-off"Per-user preferences (default on for the cell-output features) live in config.json as enable_explain_error, enable_output_followup, enable_output_toolbar.
List-shaped denylists (LLM providers, built-in tools, coding-agent launcher tiles) use traitlets rather than *_POLICY env vars. See docs/admin-guide.md for the disabled_providers, disabled_tools, and disabled_coding_agent_launchers recipes.
Value-presence locks for non-boolean settings: setting the env var to a non-empty value pins the control to that value and disables it. Empty/unset = user-choice.
| Env var | Pins |
|---|---|
NBI_CHAT_MODEL_PROVIDER | General → Chat model → Provider |
NBI_CHAT_MODEL_ID | General → Chat model → Model |
NBI_INLINE_COMPLETION_MODEL_PROVIDER | General → Auto-complete model → Provider |
NBI_INLINE_COMPLETION_MODEL_ID | General → Auto-complete model → Model |
NBI_CLAUDE_CHAT_MODEL | Claude → Chat model (and Inline chat model, unless pinned separately) |
NBI_CLAUDE_INLINE_CHAT_MODEL | Claude → Chat model → Inline chat model |
NBI_CLAUDE_INLINE_COMPLETION_MODEL | Claude → Auto-complete model |
ANTHROPIC_API_KEY | Claude → API Key (input is locked + blanked; the SDK reads the env directly) |
ANTHROPIC_BASE_URL | Claude → Base URL |
NBI_ACP_CHAT_MODEL | ACP → Chat model |
OPENAI_API_KEY | ACP → API Key (input is locked + blanked; the agent reads the env directly) |
OPENAI_BASE_URL | ACP → Base URL |
Inline chat calls the Anthropic API directly while the other Claude modes go through the Claude Code CLI, so a deployment can need a different model id for each. It uses inline_chat_model, falling back to chat_model when that is unset. NBI_CLAUDE_CHAT_MODEL therefore governs both surfaces: while it is set, a stored inline_chat_model is ignored and its control is disabled. Set NBI_CLAUDE_INLINE_CHAT_MODEL as well to pin inline chat to a different model — it takes precedence and re-enables nothing, so both surfaces stay administrator-controlled.
NBI_ACP_AGENT_COMMAND (not a settings lock) overrides the adapter command NBI launches for the selected ACP agent type; the value is shell-split. Leave it unset to use the agent's registered default command.
Provider IDs: github-copilot, openai-compatible, litellm-compatible, ollama, none. The *_MODEL_ID value is whatever the chosen provider exposes (e.g. gpt-4o, llama3:latest). Claude model IDs are the literal IDs from the Anthropic API (e.g. claude-opus-4-7, claude-sonnet-4-6); empty string = "Default (recommended)"; NBI_CLAUDE_INLINE_COMPLETION_MODEL also accepts none (no inline completion in Claude mode) or inherit (use the General-tab Auto-complete model).
Upload tunables govern the shared upload endpoint used by both chat-sidebar file attachments and terminal drag-drop:
| Env var | Default | Behavior |
|---|---|---|
NBI_UPLOAD_MAX_MB | 50 | Per-file size cap in megabytes. Requests over the cap return HTTP 413. Set to 0 to disable the cap entirely. |
NBI_UPLOAD_RETENTION_HOURS | 24 | How long staged uploads survive in the temp directory before the next upload sweeps them. Set to 0 to keep only the atexit purge (uploads survive the session). |
The same values are also configurable via the upload_max_mb and upload_retention_hours traitlets on NotebookIntelligence.
NBI can remember your GitHub Copilot login so you don't have to sign in again after a JupyterLab or system restart.
Caution
If you enable this, NBI encrypts the token and stores it in ~/.jupyter/nbi/user-data.json. Never share this file. The encryption uses a default password unless you set NBI_GH_ACCESS_TOKEN_PASSWORD to a custom value — on shared or multi-tenant systems, set a custom password before enabling this option.
NBI_GH_ACCESS_TOKEN_PASSWORD=my_custom_passwordTo enable, check Remember my GitHub Copilot access token in the Settings dialog.
If the stored token fails to authenticate (expired, revoked, password mismatch), NBI prompts you to sign in again.
These tools are available in Agent mode and to MCP-enabled chats.
| Tool | What it does |
|---|---|
Notebook Edit (nbi-notebook-edit) | Edit notebooks via the JupyterLab notebook editor. |
Notebook Execute (nbi-notebook-execute) | Run notebooks in the JupyterLab UI. |
Python File Edit (nbi-python-file-edit) | Edit Python files via the JupyterLab file editor. |
File Edit (nbi-file-edit) | Edit files in the Jupyter root directory. |
File Read (nbi-file-read) | Read files in the Jupyter root directory. |
Command Execute (nbi-command-execute) | Execute shell commands using the embedded terminal in Agent UI or JupyterLab terminal. |
In multi-tenant deployments, nbi-command-execute and nbi-file-edit are effectively arbitrary code execution as the user. See docs/admin-guide.md for guidance on disabling them.
NBI integrates with MCP servers. It supports both stdio and Streamable HTTP transports. Tools and prompts are supported; resources are not. A server's prompts appear in the chat input's slash-command autocomplete as /mcp:<server>:<prompt>.
Add MCP servers by editing ~/.jupyter/nbi/mcp.json. An environment-wide base file at <env-prefix>/share/jupyter/nbi/mcp.json is also supported.
Note
MCP requires an LLM model with tool-calling capability. All GitHub Copilot models in NBI support this. For other providers, choose a tool-calling-capable model.
Caution
Most MCP servers run on the same machine as JupyterLab and can make irreversible changes or access private data. Only install MCP servers from trusted sources.
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"/Users/mbektas/mcp-test"
]
}
}
}For stdio servers you can pass extra environment variables under env:
"mcpServers": {
"servername": {
"command": "",
"args": [],
"env": {
"ENV_VAR_NAME": "ENV_VAR_VALUE"
}
}
}For Streamable HTTP servers you can also specify request headers:
"mcpServers": {
"remoteservername": {
"url": "http://127.0.0.1:8080/mcp",
"headers": {
"Authorization": "Bearer mysecrettoken"
}
}
}To temporarily disable a configured server without removing it, set "disabled": true:
"mcpServers": {
"servername2": {
"command": "",
"args": [],
"disabled": true
}
}NBI's ruleset system lets you define guidelines and best practices that get injected into AI prompts automatically — for consistent coding standards, project conventions, or domain knowledge. Rules are markdown files in ~/.jupyter/nbi/rules/ and can scope by file pattern, kernel, directory, or chat mode.
A two-line example:
---priority: 10---- Always use type hints in Python functions.
- Add docstrings to all public functions.For full details (frontmatter reference, mode-specific rules, auto-reload), see docs/rulesets.md.
The Settings panel exposes a top-level Skills tab for managing the skills that Claude can invoke. Skills are stored under ~/.claude/skills/ (user) or <project>/.claude/skills/ (project). You can create and edit skills inline, duplicate, rename, delete (with undo), or import from a public GitHub repo. The tab is visible in any mode; when Claude mode is off, a hint banner notes that skills only take effect inside Claude sessions (handy for authoring skills now and using them when you turn Claude mode on later).
For organization-wide deployments, your admin can ship a curated manifest of skills and keep them in sync by setting NBI_SKILLS_MANIFEST. Skills installed this way are marked Managed and are read-only in the UI. Admins who want to keep managed skills but disallow ad-hoc GitHub imports use NBI_ALLOW_GITHUB_SKILL_IMPORT=false.
For full details, see docs/skills.md.
When Claude mode is enabled and the Claude CLI is available, the Settings panel exposes an MCP Servers tab that manages the user, project, and local-scope MCP entries Claude Code reads from ~/.claude.json and the project's .mcp.json. This is a different tab from NBI's own MCP Servers tab (which manages the servers used by the non-Claude chat path); the two never appear together, and the Settings dialog shows whichever one applies to your current mode.
Reads come from Claude's JSON config files directly. Writes (add and remove) shell out to claude mcp add and claude mcp remove so Claude remains the source of truth for any side effects (project-trust prompts, OAuth bookkeeping). Each server entry can be toggled on or off per workspace without removing it, and the Add MCP server dialog accepts a JSON-paste path that takes a Claude / Cursor / VS Code MCP config blob and pre-fills the form after validating the shape. Admins can lock the tab with NBI_CLAUDE_MCP_MANAGEMENT_POLICY=force-off.
When Claude mode is enabled and the Claude CLI is available, the Settings panel exposes a Plugins tab wrapping claude plugin for install, uninstall, enable, disable, and marketplace add (for example: add a marketplace from a GitHub repo, then install plugins it publishes). Each installed plugin's description, author, version, and source render inline, and the tab surfaces a per-plugin Update button when a newer version is available upstream. A marketplace picker lets you browse the configured marketplaces and install plugins directly from there. Marketplaces hosted on GitHub reuse the same GITHUB_TOKEN / GH_TOKEN / gh auth token precedence as Skills imports; the token is passed via env to the subprocess and never appears in argv or DEBUG logs. See Anthropic's plugin docs for what a plugin is and how marketplaces work.
Admins can lock the entire tab with NBI_CLAUDE_PLUGINS_MANAGEMENT_POLICY=force-off, or keep the tab and block only GitHub-sourced marketplaces with NBI_ALLOW_GITHUB_PLUGIN_IMPORT=false.
Enable thumbs-up/down feedback on AI responses by setting:
c.NotebookIntelligence.enable_chat_feedback=True…or via CLI:
jupyter lab --NotebookIntelligence.enable_chat_feedback=trueThe feedback fires an in-process telemetry event. Nothing leaves the process by default. See the admin guide for the full set of telemetry events and how to wire them into your observability stack.
The thumbs buttons reveal on hover by default. To keep them always visible, enable:
c.NotebookIntelligence.enable_chat_feedback_always_visible=True
NBI Settings opens on a Status card that answers one question: is this deployment configured to work, and if not, which specific piece is missing? Each check reports ok, warn, or blocked, and anything that is not ok carries the next action rather than just the failure ("the claude CLI was not found" plus "install it, or set NBI_CLAUDE_CLI_PATH and restart").
It runs on open and costs nothing: it resolves your provider and model, checks the model list is reachable and that your selected model is one the endpoint actually serves, and for the agent modes checks that the CLI answers --version and that credentials are present. A missing key is a warning rather than an error, because the Claude CLI and ACP agents can hold their own subscription logins that NBI cannot see.
Test the endpoint sends one short request to your configured model, after asking first. It is the only way to catch two failures nothing cheaper can see: a gateway that returns 200s but buffers instead of streaming, and a proxy that strips the tools field so agent mode silently never calls a tool. No model output is kept, so the result is safe to paste into a support ticket.
That button is the only readiness check that bills anything, so it has an off switch: NBI_READINESS_LIVE_CHECK=off (traitlet readiness_live_check_allowed, default on) refuses it while leaving every other check available. One live test runs at a time per server; a second concurrent request gets a 429 rather than spending tokens.
An opt-in mode for answering "where did this turn's time go," aimed at deployments where NBI feels slower than it should: internal LLM gateways, network home directories, TLS-intercepting proxies. Off by default, and when off the cost is a single boolean check per instrumentation site.
Turning it on. NBI Settings, Performance tab, Enabled. It takes effect immediately with no restart.
Reading a turn. Every chat turn records a phase timeline (rule and context preparation, agent connect and spawn, time to first token, streaming with stall events, tool calls, time spent waiting on you) plus token counts and the SDK-reported API duration. The table shows the recent turns, one row each, with a Verdict column naming the phase that dominated. The comparison that matters most is Active against API ms: close together means the time is in the gateway or the model, far apart means it is local. Expanding a row shows that turn's spans as proportional bars with their attributes. Each turn is also summarized as one perf turn ... INFO line in the Jupyter server log, so headless installs get the same signal with no UI.
Reading the environment.Run probe measures the machine instead of a turn: filesystem latency and throughput for the config and ~/.claude directories with filesystem type and mount options, interpreter and CLI cold-start cost, and process and host contention. Results are grouped and banded, so a slow network filesystem or a throttled container is called out in a sentence rather than left in a JSON blob. An opt-in network check adds connections to your configured endpoint to time DNS, TCP, and TLS and to capture the presented certificate chain, which is what makes TLS interception visible. It asks for confirmation first and names the host it will contact.
Privacy. Recorded: durations, counts, byte sizes, status enums, token counts, and tool, server, model, and provider names. Never recorded: prompt or response text, file contents, absolute paths, environment variable values, API keys, exception messages, or hostnames. This is enforced by a fixed per-attribute allowlist rather than by convention, and the default redacted detail level additionally hashes file basenames and model, tool, and server names.
Collecting across sessions. Turns can optionally be appended as JSON Lines to <nbi-dir>/perf/, with size- and age-capped retention. On network home directories, point the log at local scratch with NBI_PERF_LOG_DIR.
Full guide, including the span and event reference, the probe thresholds, worked diagnoses for a slow gateway, a slow home directory, and an intercepting proxy, and the JSONL schema: docs/performance-diagnostics.md.
| Environment variable | Effect |
|---|---|
NBI_PERF_DIAGNOSTICS | Locks the enabled state on or off (value lock; the UI shows the toggle as locked). |
NBI_PERF_DIAGNOSTICS_POLICY | user-choice (default), force-off, or force-on (fleet-wide collection with redacted attributes enforced). |
NBI_PERF_PROBE_NETWORK | off disables just the probe's network check while leaving the rest available. |
NBI_PERF_LOG_DIR | Redirects the JSONL perf log directory. |
docs/admin-guide.md— deployment, env vars, security model, air-gap, multi-tenancy.docs/skills.md— Claude Skills management and the org-manifest reconciler.docs/rulesets.md— ruleset frontmatter and discovery.docs/troubleshooting.md— common problems with copy-pasteable fixes.docs/performance-diagnostics.md: turn timelines, the environment probe, and how to read both.PRIVACY.md— what NBI sends to which provider, and the egress allowlist.SECURITY.md— how to report a vulnerability.CHANGELOG.md— release history.CONTRIBUTING.md— building NBI from source. Skip this if you just want to use NBI.
- Introducing Notebook Intelligence!
- Building AI Extensions for JupyterLab
- Building AI Agents for JupyterLab
- Notebook Intelligence now supports any LLM Provider and AI Model!
NBI 5.x is stable. New features land in minor releases (5.1, 5.2, …); breaking changes are reserved for the next major (6.x) and will be announced in the changelog. Upgrading from 4.x? See the 5.0.0 migration note for the fastmcp → mcp dependency swap, the new path sandboxes, and the workspace-file-attach behavior change.
Licensed under GPL-3.0.