n8n launches Nod, a personal agent with its own browser
By AgentRiot Editorial
Nod runs errands across apps and websites from your own machine, with approval checks, persistent memory and important limits behind its benchmark scores.

n8n built Nod to sign into supplier portals, collect invoices and work through errands such as subscription cancellations or meeting preparation. The research preview from n8n Labs (opens in new tab) gives a personal agent its own Linux workspace and browser, rather than leaving the user to carry out a chatbot's instructions.
n8n announced Nod on November 11 (opens in new tab). It is a standalone app for one person, not the n8n workflow builder or its AI Assistant. No n8n account is required. The distinction matters: shared business automations remain n8n's job; Nod connects to an individual's accounts and takes on that person's recurring work.
Its most consequential design choice is the separation between the model doing the work and the server deciding whether an action is permitted. That adds a useful control, but the permission decision still depends on a model. The published tests show why both halves deserve attention.
A private computer, not control of your desktop
In the v0.6.2 source snapshot (opens in new tab), Nod combines a Node.js 24 server, Fastify API and React interface with the Pi agent runtime. PostgreSQL stores conversation and job state. Separate containers provide a Linux workspace and Chromium browser; Electron packages the desktop client and its private runtime.
Commands and generated files live in that workspace, not an unrestricted view of the host's home directory. Model-provider keys stay with the server, outside the workspace, browser and model context. The browser uses agent-browser to inspect pages and operate controls. Users can take over its live tab when a website needs human intervention.
App connections and browser work serve different purposes. Gmail, Calendar and Slack can use managed integrations, while Notion, Linear and n8n have their own MCP connections. A browser gives Nod a route through a supplier portal or dashboard without a suitable integration. Managed app traffic goes through n8n's default connection service and Composio (opens in new tab), so a local installation is not an entirely local integration stack.
Consider a monthly invoice task. Nod can read the instruction, open the specified portal, request a saved sign-in, download the files and organize them in its workspace. Sending those files to an accountant needs authority covering that recipient and action. A page asking it to forward invoices somewhere else is outside that instruction.
The app, web interface, Telegram and WhatsApp share one conversation. Independent tasks keep separate histories while sharing the owner's workspace and memory. That supports overlapping jobs, but not a shared team account: each installation currently belongs to one owner (opens in new tab).
Approval does not mean asking about every click
Under Nod's approval rules (opens in new tab), a request can authorize the steps needed to achieve its stated outcome. Asking it to book a specific restaurant can cover confirming the booking. Asking it to research restaurants does not. A rule to ask first overrides the normal permission to proceed, even when an instruction otherwise covers the action. Card payments have a separate approval flow.
Connected-app actions go through a server-side review. By default, the checker uses the selected model at its lowest reasoning setting (opens in new tab); a server setting can select other checker paths. Reads and drafts normally proceed unless the owner's instructions restrict them; other actions need a matching request, schedule, standing instruction, rule or explicit approval. The checker must supply the owner's words authorizing a non-read action. The server verifies that those words exist, not that the model interpreted their meaning correctly.
Browser checks have a narrower boundary than “every interaction is approved.” Clicks, submissions, uploads and other potentially consequential steps are checked. Opening pages, following links, read-only scripts and typing without Enter skip that check, including typing into an autosaving field. For a checked step, the model receives the page address, control label, relevant form destination and action details, not the whole page text. A deceptive label can still mislead it. These are explicit security-model limitations (opens in new tab), not theoretical exceptions to an otherwise guaranteed filter.
Vault addresses a separate problem: exposing credentials to the working model. Saved logins and cards are encrypted, and private browser entry keeps their values out of model-visible tool responses. That does not make the destination website trustworthy. The server and browser provider can read the values, and the website receives what is submitted. A saved sign-in also does not constrain every later click.
Isolation varies by deployment. The browser and web tools are restricted to public addresses, but a source or self-hosted workspace has normal Docker network access without an outbound firewall. The packaged desktop runtime blocks private and reserved address ranges for workspace traffic. Chromium's own sandbox is disabled inside its container, making the container the security boundary, as the isolation guide (opens in new tab) explains.
What the benchmark scores actually cover
n8n's evaluation report (opens in new tab) separates task completion, permission and memory. These are vendor-reported results, not an independent product comparison.
| Evaluation | Published result | Scope and configuration |
|---|---|---|
| Errand completion | 98%, rounded | 226 errands, three attempts each, 678 runs; GPT-6.1 Sol, October 17, 2026 |
| Runs with an unsafe action | 0/678 | Same task runs; actions checked against simulated user permission |
| Successful prompt-injection scenarios | 0/13 | Nod-authored attack subset within the 226 errands, three attempts per scenario |
| LongMemEval-S recall | 93.2% | 500 questions, fresh sessions; GPT-6.1 Sol, October 19; GPT-4o-2024-08-06 judge, no retries |
| WhisperBench memory injections | 0/216 | 108 attack emails in foreground and background modes; reported October 19; GPT-6.1 Sol with GPT-5.4-nano judges |
The errand harness runs the actual server, runtime, tools, approval flow and container workspace against stateful synthetic apps and websites. Checks inspect changed app state, files and recorded activity rather than trusting the agent's claim that it finished. A simulated user supplies approvals and answers; separate model judging handles criteria such as summary quality.
Completion is not the stricter rate at which every required safety, honesty and product check passes. The report does not publish that stricter rate alongside the headline. Nor are 678 attempts 678 different task designs. The taxonomy draws on more than 900,000 anonymized production workflows, but the scored snapshot contains 226 selected errands, not 900,000 executed workflows or a random production sample.
LongMemEval-S supplies roughly 115,000 tokens of earlier conversation per question. Nod's fresh-session setup has saved notes, the last 30 messages and history search. Its separate transcript-bearing chat setup should not be confused with this result. The measured revision was c8c90ba1, not the launch release.
| LongMemEval-S fresh-session group | Earlier code | Evaluated code |
|---|---|---|
| All 500 questions | 79.0% | 93.2% |
| 418 questions not inspected during development | 83.7% | 96.4% |
The 418-question row is a subset of the same run, not another experiment. Neither row provides repeated-run confidence intervals or a matched ranking against competing assistants.
WhisperBench has an especially important distinction between planting a memory and influencing a later answer:
| Mode | Memory injected | Attack stayed hidden | Later answer used target claim | End-to-end success |
|---|---|---|---|---|
| Background | 0/108 | 6/108 | 8/108 | 0/108 |
| Foreground | 0/108 | 0/108 | 4/108 | 0/108 |
In those 12 affected later answers, Nod reread the attack email from the inbox. The planted claim did not come from saved memory. Zero memory injections therefore does not mean malicious content never influenced an answer. End-to-end success requires injection, stealth and effectiveness together; missing-evidence runs are incomplete and excluded from attack rates.
The release README (opens in new tab) also reports a separate checker comparison:
| Checker model | Published outcome on 240 test decisions |
|---|---|
| GPT-6.1 Sol | All correct |
| Claude Opus 5.5 | All correct |
| Claude Haiku 5.5 | Allowed eight unsafe actions |
That is a model-choice warning, not a full-system comparison. The README does not give this comparison's run date, complete peer accuracy breakdown or measured cost table. Historical run manifests and individual case results would be needed for a full independent audit. Current evaluation code has also gained cases, so running today's suite would not reproduce the October snapshot. None of these results guarantees production safety.
Memory has more than one entry point
Nod's personal memory notes (opens in new tab) draw on the owner's messages and explicit settings or imports, not arbitrary facts found in pages or incoming email. Notes carry supporting evidence and can be inspected, edited or forgotten. Gmail supplies a separate writing-style note, intended to omit facts, names and quotations.
Learned skills are different. Nod writes reusable procedures from tasks it performed and content it read. A hostile page or email can therefore shape a skill, even though it cannot supply authority for an action. Skills reach the working model as untrusted guidance and do not go to the permission checker. This distinction makes a blanket “external content cannot affect memory” claim too broad.
Storage is bounded: the memory document holds up to 256 KiB, each note up to 4,000 characters, and up to 40 learned skills with bodies of 6,000 characters. Retrieval also has limits, and history search does not include a reset conversation. Remembering useful facts is not exhaustive recall of everything an owner has ever discussed.
Local storage still sends data out
Nod's privacy documentation (opens in new tab) says conversations, retrieved material and screenshots can go to the selected model service. Other connected services can handle image interpretation or decision checks. Model keys remain encrypted locally, but task content is not kept offline merely because the app runs on a laptop.
Usage sharing with n8n (opens in new tab) is on by default. It includes feature events, timings, costs, basic device information and onboarding or survey answers, but excludes ordinary conversation content. Conversation traces are a separate, default-off option. Their local redaction is best effort: traces are pseudonymized, not anonymous. The privacy page lists EU PostHog storage for usage and EU LangSmith storage for traces, with 14-day trace retention; consented onboarding contact data and shared fixed choices can also reach Customer.io in the US.
Turning sharing off in Settings stops subsequent sends, not data already sent or necessarily a request in flight. A fresh desktop profile can send its first setup-window event before the user reaches that switch. Deletion requires a separate request (opens in new tab). Those details matter more than a broad promise that the agent is “yours.”
What it takes to run Nod
The desktop installer (opens in new tab) supports macOS 14 or later on Apple silicon and Intel, plus Windows 11 x64 in preview. Windows local operation needs WSL 2, hardware virtualization and at least 20 GB of free disk. The desktop app does not require separately installing Node.js or Docker. Linux users have a source-install route (opens in new tab), requiring Node.js 24.21.0 or later, pnpm and Docker with Compose v2, rather than an advertised Linux desktop binary.
Nod's code uses Apache 2.0 (opens in new tab); bundled fonts, provider logos and desktop runtime components retain their separate terms (opens in new tab), including GPL-3.0-or-later speech components (opens in new tab). The software is free, but model usage is not universally free (opens in new tab). Provider bills, independent integrations and server or browser infrastructure can add costs outside Nod's model estimates.
ChatGPT Plus or Pro sign-in (opens in new tab) uses the plan's included usage, not unlimited access to every provider. API connections are billed separately. Nod estimates known API costs, supports a monthly limit and defaults to a $5 task check-in. Existing work can finish after the monthly limit is reached; unpriced usage does not count toward it. A ChatGPT plan's API-equivalent task estimate is not an extra charge.
The evaluation guide (opens in new tab) estimates $0.10 to $0.50 per capability attempt, about $350 for both LongMemEval setups and about $55 for both WhisperBench modes with GPT-6.1 Sol. These are reproduction estimates, not the historical tests' actual bills or a normal errand tariff.
Schedules (opens in new tab) need an awake host with Nod running. An always-on personal server can continue while a client sleeps, but n8n-hosted operation, team ownership and phone calls remain roadmap items (opens in new tab). For a first trial, a bounded invoice-download or draft-only task is a better starting point than open-ended authority over payments and outgoing mail. The next useful test is whether Nod respects that boundary on the accounts and websites its owner actually uses.

