Heptabase Review 2026: Is This Visual AI Knowledge Base Worth It?

Heptabase released version 1.102.0 on August 4, 2026. The headline change is not a new note type or another isolated model option. AI chat can now use the active view as context, including selected objects on a whiteboard and the current page of a PDF. Users can also drag images directly into the chat composer, while failed AI Agent runs preserve partial output and offer a retry route. Together, these changes reduce the distance between looking at research and asking the agent to work with the exact material in front of you.

Active-view context matters because visual knowledge systems contain meaning that ordinary text search can miss. A user may place one group of evidence beside a competing explanation, isolate unanswered questions in a section, or open the precise PDF page that contains a method or limitation. Heptabase can now pass that immediate working context into AI chat more directly. The release also improves Agent performance when listing cards, tags, and whiteboards, and improves reliability when searching cards or editing database properties. These operational changes make the agent more useful for real knowledge work than a model that only responds to copied text.

The safest interpretation is still narrower than saying the AI automatically understands an entire visual workspace. Active-view context covers the supported current view and selection. It does not guarantee that every distant card, hidden relationship, visual convention, embedded file, or unstated reason behind a layout is understood. The user must still make the question explicit, select the correct objects, check which items the agent accessed, and verify the result against the source. Spatial proximity can suggest a relationship, but it is not the same as a written claim.

Image input expands the research surface. A diagram, scanned excerpt, chart, photograph, screenshot, or visual note can be dragged into the composer. Model-specific limits still apply. Heptabase's Help Center lists different maximum file sizes, total request sizes, and image counts for OpenAI, Anthropic, and Gemini routes. Mentioning a card also does not automatically make the model read every image embedded inside that card. Users should add the relevant image through a supported route and reduce the request when a provider limit is reached.

Reliability improvements are equally important. Agentic editing creates a different failure risk from ordinary chat because a run can search, reason, and change persistent content. Preserving partial output helps users see what happened before a failure and retry without starting from a blank response. It does not prove that the partial work is complete, consistent, or safe to accept. Before retrying, inspect which cards, boards, tags, or properties were already changed. Use version history, snapshots, and manual backups as recovery controls, not as an excuse to allow broad unreviewed writes.

The August release follows a concentrated July expansion. Version 1.99.0 introduced skills, Space Instructions, AI chat tabs, basic post-limit chat for Premium users, and the ability for Agent to organize wider parts of the knowledge base. Version 1.100.0 added OpenRouter and several open-source model choices. Version 1.101.1 added a route for using ChatGPT Codex inside Heptabase and made the composer show whether the next request would use a ChatGPT plan, a personal API key, or Heptabase credits. Version 1.101.2 added control over which models appear in model pickers.

That funding label is a valuable detail because Heptabase now supports several cost paths. Pro, Premium, and Premium+ include 100, 1,800, and 8,100 monthly credits. Those credits reset and do not roll over. Premium plans can continue with designated Basic models after the allowance ends, but that mode can read, search, and answer only. It cannot create or edit content. On-demand usage is billed after use, with charges triggered around a $10 running balance or at the billing-cycle end. A personal API key creates direct provider charges, and a ChatGPT connection depends on the user's separate plan and eligibility.

Buyers should therefore measure the cost of completed outcomes, not the number of credits printed on a plan. A simple question, a long PDF analysis, a transcription, a Tutor lesson, and an agentic database reorganization may consume very different resources. During the trial, log the model, route, credits, external provider charges, result quality, and review time for each representative task. Premium+ is only better value when the extra allowance is used productively, not when broad context or repeated retries consume it without improving the work.

The privacy boundary also deserves attention. Space search is currently on by default in AI Agent conversations. Heptabase says its retrieval system can search all cards and whiteboards in the enabled Space, then normally sends a smaller group of relevant items to the model. The user cannot exclude one sensitive card while leaving all other content in that Space searchable. For confidential work, switch Space search off and add only approved objects manually. Heptabase's terms also say its AI Services must not be used to process sensitive personal data.

Standard integrated AI requests may send content to OpenAI, Google, or Anthropic. Heptabase says User Content is not used to train its models and that its provider agreements prevent training general foundation models on that data, although limited provider retention for abuse monitoring can occur. A user-provided API key has a different legal and privacy route. Heptabase's negotiated protections do not apply, so the user's provider account settings, agreement, retention, residency, and training controls become decisive.

Version 1.102.0 also improves the connection between the desktop knowledge base and coding agents. Official CLI skills can let agents read cards and tags through deep links, while earlier CLI releases added cards, journals, whiteboards, tag properties, parsed PDF pages, transcripts, and AI Tutor content. This is separate from Heptabase MCP. MCP works with cloud-synced data through OAuth and supported remote tools. The CLI works through the running desktop app and local database. Heptabase explicitly warns users not to bypass official CLI commands or modify application data through database files, caches, or internal endpoints.

The best rollout is deliberately narrow. Start with one active research board, one PDF, one small tag database, and a recoverable test card. Ask the agent to explain what it can see before allowing edits. Require citations or object links for factual claims. Review every proposed structural change, check version history after writes, and export a full backup that includes attachments. If MCP or CLI is connected, begin with read-only tasks and verify the exact Space, account, and permissions.

WhatAI's take: Heptabase is becoming one of the more integrated visual AI knowledge workspaces. Active-view context, skills, Agent editing, Tutor, MCP, CLI, and multiple model routes can create a powerful loop between reading, thinking, structuring, and acting. The tradeoff is complexity. More context and more write access increase the importance of scoping, cost tracking, source verification, privacy review, and recovery. Heptabase is most convincing when the agent helps a user understand a difficult subject while the human remains clearly in control of evidence and permanent changes.

WhatAI Decision Box

Heptabase combines connected note cards, visual whiteboards, tag databases, PDFs, web content, highlights, journals, collaboration, and native AI in one research and learning workspace. Its strongest use case is making sense of complex material over time, not simply writing isolated notes.

Heptabase is a paid product with a seven-day trial. Pro starts at $8.99 per month on annual billing, Premium at $17.99, and Premium+ at $53.99. Monthly billing costs $11.99, $23.99, and $71.99. Included AI credits reset each cycle, while on-demand usage and personal API keys can add separate costs.

AI Agent can search an enabled Space, inspect selected cards, whiteboards, databases, PDFs, images, and transcripts, then answer questions or edit supported content. Space search is on by default and cannot exclude one sensitive card, so privacy-conscious users should disable it and attach only approved context.

Use Cases

Pricing varies by plan and region — see current pricing.

Plan features change — last updated: 2026-08-09.

Heptabase Pros & Cons

How to Get Results with Heptabase: Step-by-Step Workflow

1. Define the learning or research outcome

   Choose one complex topic, project, literature review, course, or decision. Define what you must understand, produce, or decide by the end of the trial.

2. Select an import and capture boundary

   Import only representative notes and sources. Connect Zotero or Readwise only if they are part of the real workflow, and avoid filling the space with an unreviewed archive.

3. Create a source whiteboard

   Place the central question, key source cards, PDFs, web cards, and an unanswered-questions section on one board. Use cards as reusable knowledge objects instead of copying the same note.

4. Read and extract evidence

   Highlight PDFs, notes, and web material. Turn only decision-relevant excerpts into cards, preserve the source and page, and separate quotations, interpretations, and unanswered claims.

5. Build a minimal tag database

   Create a small schema for source type, status, theme, confidence, and next action. Add table or Kanban views only when they support an actual review or decision.

6. Connect the visual argument

   Group related cards, add labeled connections, use sections for competing explanations, and create sub-whiteboards when one cluster becomes a separate problem.

7. Configure AI deliberately

   Choose the model and billing route, set Space Instructions, disable Space search for sensitive work, attach approved context, and require references or visible evidence for factual conclusions.

8. Use Agent or Tutor on a bounded task

   Ask AI to identify gaps, compare evidence, propose a syllabus, or reorganize selected material. Review every source, edit, property change, and generated artifact before accepting it.

9. Test collaboration and agent connections

   If needed, share one board with least-privilege permissions and test MCP or CLI on read-only tasks before allowing writes. Confirm which account, Space, and data route each tool can access.

10. Export, restore, and measure

    Create a manual full backup, export notes, restore a test version, and inspect migration output. Continue only if the system improves understanding and retrieval enough to justify subscription and AI costs.

Heptabase Gotchas and Limits to Know Before You Start

• Heptabase has a seven-day trial but no directly accessible permanent free personal workspace.
• The restricted free plan exists only for people invited to collaborate by a paid user and omits personal knowledge features and AI.
• The trial ends after seven days or when its AI credits are exhausted, whichever comes first.
• Annual prices are lower monthly equivalents and require an annual payment of $107.88, $215.88, or $647.88.
• Included AI credits reset on each billing cycle and do not carry over.
• AI credit consumption varies by model and feature, so the number of useful tasks cannot be inferred from the credit total alone.
• On-demand usage is postpaid, charges when the running total reaches about $10 or at cycle end, and a final request can push the charge slightly above the monthly cap.
• Premium and Premium+ basic post-limit chat can read, search, and answer, but cannot create or edit content.
• Bring-your-own API keys create separate provider costs and apply the user's direct provider terms rather than Heptabase's negotiated protections.
• A connected ChatGPT plan is a separate billing route and depends on model, plan, workspace, and feature eligibility.
• Space search is on by default in AI Agent and can search all cards and whiteboards in the enabled Space.
• Users cannot exclude one sensitive card or whiteboard while leaving the rest of the same Space searchable.
• Heptabase's terms prohibit processing sensitive personal data through its AI Services.
• Standard integrated AI can transmit text, audio, video, and other content to OpenAI, Google, or Anthropic for processing.
• Third-party AI providers may retain request data temporarily for abuse monitoring even when it is excluded from general model training.
• Cloud data uses server-side AES-256 encryption at rest and TLS in transit, not a documented zero-knowledge end-to-end encryption design.
• Cloud data and Elastic web search are documented as operating in AWS us-west-1, which may not meet every residency requirement.
• The published Privacy Policy was last updated in December 2021, while newer AI processing details are in the December 2025 Terms.
• Cloud sync is enabled by default, and deactivating active sync is no longer supported through the normal product route.
• Automatic cloud backups exclude attachments, so manual full backups are required for complete recovery coverage.
• Version history lasts sixty days and does not replace an independent long-term backup.
• Whiteboard restoration recovers layout and objects but not the content history of every card placed on the board.
• Heptabase allows files up to 2 GB and no stated overall storage cap, but desktop uploads consume local device storage.
• Premium is the tier that publicly advertises unlimited PDF uploads and unlimited parsing; do not assume identical Pro PDF allowances.
• Zotero sync is one-way, must be triggered manually, skips attachments over 100 MB, and its official FAQ lists first-version mobile and source-card limitations.
• Readwise sync is not fully bidirectional, and edits made in Heptabase can stop later Readwise changes from overwriting that highlight.
• Notion, Roam Research, and Logseq imports are partial or beta, and block-level references or complex databases may not transfer.
• Markdown export does not guarantee that whiteboard layout, tag databases, properties, public links, permissions, or every attachment will behave identically elsewhere.
• Collaborators see only explicitly shared boards and associated cards, but permission behavior differs across cards, journals, deletion, history, and Card Library actions.
• Embedded cards may be invisible to collaborators unless the cards are explicitly placed on the shared whiteboard.
• The CLI must use official commands through the running desktop app; direct access to local database files can corrupt data.
• The CLI can edit values in an existing tag database but cannot create or redesign the tag database schema through current supported commands.
• MCP accesses cloud-synced data, while CLI accesses the running local app, so availability, permissions, and security boundaries differ.
• AI answers, source selection, summaries, transcriptions, database edits, whiteboard changes, and tutoring material can be incomplete or wrong.
• The Help Center describes refunds case by case, while the governing terms say post-September 2023 fees are non-refundable, so no refund should be assumed.
• The affiliate program requires application and approval before WhatAI can use a commission-tracking link.

Which Heptabase Feature Fits Your Use Case

Starter Prompts for Heptabase

What is Heptabase?

Heptabase is a visual knowledge and research workspace. It combines cards, whiteboards, tags, structured databases, journals, tasks, highlights, PDFs, web and media cards, collaboration, native AI, mobile and desktop apps, cloud sync, MCP, and a CLI. It is designed for understanding complex topics across connected sources.

Is Heptabase free?

Heptabase does not offer a directly accessible permanent free personal workspace. New buyers can use a seven-day trial. Non-subscribers invited by a paid owner can receive restricted free collaborator access to shared whiteboards, but they do not receive a personal journal, Card Library, tag database, tasks, highlights, or AI features.

How much does Heptabase cost?

Standard pricing is $11.99 monthly or $107.88 yearly for Pro, $23.99 monthly or $215.88 yearly for Premium, and $71.99 monthly or $647.88 yearly for Premium+. The annual equivalents are $8.99, $17.99, and $53.99 per month. Taxes and separate AI usage can increase the final cost.

What is included in the Heptabase trial?

The trial lasts for up to seven days or until the trial AI credits are exhausted, whichever happens first. It includes the Premium-level AI credit allowance plus unlimited PDF uploads and parsing during the trial. The selected paid subscription begins after the trial unless it is cancelled.

How do Heptabase AI credits work?

Pro includes 100 credits per month, Premium 1,800, and Premium+ 8,100. Credits can be consumed by AI Agent, video transcripts, Insight Generator, and voice notes. They reset each billing cycle and do not carry over. Cost per action varies with the feature and model.

What happens after Heptabase AI credits run out?

Premium and Premium+ users can continue using designated Basic models for read, search, and answer chat without additional on-demand usage. That post-limit mode cannot create or edit knowledge-base content. Users can also enable on-demand billing or connect their own supported API key. Pro users normally need to upgrade, use on-demand billing where available, or bring a key.

Does Heptabase support ChatGPT, Claude, Gemini, and OpenRouter?

Yes, but access and billing differ. Included plans support model sets based on tier, Premium tiers include providers such as OpenAI, Google, and Anthropic, and Heptabase also supports certain OpenRouter models. Users can connect personal OpenAI or Anthropic keys, and supported ChatGPT plan connections can fund eligible requests separately.

What is Heptabase AI Tutor?

AI Tutor is a guided learning mode that can use selected sources and workspace content to build and refine learning material, lessons, and a syllabus. It is included with Premium and Premium+, while Pro users may need on-demand usage. Generated explanations, sources, and edits still require human review.

What can Heptabase AI Agent do?

AI Agent can search and read supported cards, whiteboards, tag databases, PDFs, images, transcripts, journals, and other selected context. It can answer questions and, where supported, create or edit cards, journals, whiteboard content, layouts, tags, and properties. Available actions can depend on the current version, model, plan, context, and billing route.

What data can Space search send to an AI model?

When Space search is enabled, Heptabase can search all cards and whiteboards in that Space. Its retrieval system normally selects around twenty relevant items and sends that subset with the conversation to the model. Users cannot currently exclude one sensitive card while keeping the rest of a Space searchable, so disable Space search for sensitive work.

Does Heptabase train AI models on my notes?

Heptabase's terms say it does not use User Content to train its AI models and has agreements preventing standard integrated providers from using the data to train general foundation models. Providers may retain data temporarily for abuse monitoring. These negotiated protections do not apply when a user connects a personal API key.

Is Heptabase end-to-end encrypted?

Heptabase documents AES-256 server-side encryption at rest and SSL/TLS in transit for cloud data, with storage in AWS us-west-1. It does not describe normal cloud sync as end-to-end encrypted. Organisations needing zero-knowledge encryption should not assume that server-side encryption provides the same boundary.

Does Heptabase work offline?

The desktop app stores data in a local database and supports offline snapshots that upload after reconnection. Cloud-dependent features, web access, cross-device sync, MCP, public links, integrated AI, provider calls, and some imports or connections require a network. Test the exact desktop and mobile workflow before relying on offline access.

Can Heptabase import notes from other apps?

Official import guidance covers Notion, Obsidian, Roam Research, and Logseq, with several routes marked beta or limited to pages and journals. Heptabase also supports .docx note-card imports and connected Zotero and Readwise workflows. Complex databases, block references, properties, layouts, and plugin behavior may not transfer completely.

Can I export my Heptabase data?

Yes. Heptabase provides Markdown export and a manual full-backup route. Automatic cloud backups preserve positions, states, and metadata but do not include attachments, so users should create manual full backups regularly. Migration testing should cover whiteboard layouts, tags, properties, references, media, and attachments rather than checking text notes alone.

How long is Heptabase version history?

Heptabase says it keeps sixty days of version history. Cards, journals, text elements, and highlights use timed snapshots, while whiteboards are saved on a separate interval. Whiteboard restoration recovers layout and objects but not card contents, which must be restored through each card's own history.

How does Heptabase collaboration work?

Owners share specific whiteboards and associated cards, not an entire private Space by default. Invitees can receive Full Access, Can Edit, or Can View permissions. Non-subscribers can join as restricted free collaborators. Permission details differ across spaces, cards, whiteboards, journals, version history, permanent deletion, and Card Library actions.

What is Heptabase MCP?

Heptabase MCP is an OAuth-based server for compatible AI assistants. Its supported tools can search cloud-synced objects, inspect whiteboards and full objects, retrieve journals, search and read PDF pages, create note cards, and append to journals. The external assistant's plan, developer settings, permissions, and model terms still apply.

What is the Heptabase CLI?

The official CLI lets compatible coding agents work through the running desktop app. It can create, read, and edit cards and journals, search the Card Library, manage tags and whiteboard membership, edit existing tag properties, read parsed PDF pages and transcripts, and inspect AI Tutor content. It is not a general hosted REST API.

Does Heptabase have an affiliate program?

Yes. Heptabase's official Help Center advertises a lifetime commission of 20 percent for approved creator partners. Applicants should already understand and preferably use Heptabase, then apply through the official Typeform. WhatAI should use a direct product link until its own application is approved and an assigned tracking link is issued.

Related Productivity Tools

Explore the Network