Unabyss

Unabyss is a portable context layer for AI tools. Instead of letting each chatbot or agent build a separate memory silo, it stores and structures a user’s personal and professional context, keeps that context updated from connected applications, and exposes selected pieces of it to external AI clients through the Model Context Protocol (MCP).¹²

The current product is aimed at people who move between tools such as ChatGPT, Claude, Cursor, Codex, Gemini CLI and other agents. The idea is simple: a founder should not have to re-explain company strategy in ChatGPT, then repeat project history in Claude, then restate coding conventions in Cursor. Unabyss tries to make that context portable.¹³

The product is more infrastructure-like than a conventional “memory app.” It includes a context store, Context Chat, structured files/documents, semantic search, integrations, AI-generated markdown Exports, a browser extension, desktop/application surfaces and an MCP endpoint with both read and write tools.⁴²

This is a desk review. AiToolMap did not connect Gmail, Drive, Slack, GitHub or LinkedIn; did not upload private context; did not configure an MCP token; did not measure retrieval precision; did not test permission leakage across credentials; and did not compare answers with versus without Unabyss context.

The strongest part of Unabyss is that it attacks a structural problem rather than adding another chat interface.

ChatGPT memory helps inside ChatGPT. Claude memory/project context helps inside Claude. Cursor and coding agents maintain their own rules/context. A user can manually maintain files, custom instructions or prompts, but those fragments become stale and inconsistent.

Unabyss positions itself above those tools. Its current homepage says users can connect sources such as GitHub, Notion, Linear, Obsidian, Slack, X/Twitter, LinkedIn, Gmail, Google Drive, Google Calendar, meeting tools and other applications, then let multiple AI clients query the same context.¹³

The integrations catalog currently lists MCP clients including Claude, Cursor, Claude Code, OpenClaw, ChatGPT, Grok, Vellum, VS Code, OpenCode, Codex, Gemini CLI and others; Perplexity is marked “Soon.” Data-source integrations include LinkedIn, website sync, Obsidian, X/Twitter, Notion, Slack, GitHub, GitLab, Linear, Jira, Gmail, Drive, Calendar, HubSpot, Pipedrive, Todoist, ClickUp, Monday.com, tl;dv, Fathom, Fireflies, Granola, Xodo Sign, DocuSign, Asana and Google Ads, with some additional sources marked “Soon.”³

That breadth makes the proposition compelling for exactly the kind of user whose context is distributed across many services.

Unabyss’s Security & Trust page says connected applications are imported using read-only scopes and that the service does not act on the user’s behalf inside those source applications.⁶

The Privacy Policy gives a concrete Google example: `gmail.readonly`, `calendar.readonly` and `drive.readonly` are used for Gmail messages/threads, calendar events and document content. It says Google user data is not used for advertising, is not sold, and is not used to train AI models.⁵

This is a sensible security boundary for a context product. Unabyss needs to read source material to build memory, but it generally does not need permission to send email, modify calendars or change source files.

Read-only does not eliminate risk. A compromised context layer can still expose highly sensitive information. But it limits one class of harm: an attacker or malfunctioning integration should not automatically inherit broad write privileges over the source applications merely because Unabyss is connected.

Users should still inspect actual OAuth scopes at connection time, because the integration catalog can evolve.

The current MCP documentation is much more detailed than the typical “we support MCP” marketing claim.²

Unabyss exposes a Streamable HTTP JSON-RPC MCP API at `https://mcp.unabyss.com`, implements MCP protocol version `2025-06-18`, and supports both static bearer tokens and OAuth-issued credentials.²

The tool surface includes natural-language `query`, deeper `agentic_query`, durable `store`, identity updates, integration discovery/connect proposals, export creation/read/list operations, skill discovery and setup helpers.²

Crucially, credentials can carry different redaction settings. The documentation says a token can exclude private information, exclude company-confidential material or block specific applications. Two credentials for the same user can therefore return different context to the same question.²

That is the right architecture for portable memory. “Share all my memory with every agent” would be convenient but unsafe. Per-credential redaction creates a path to different trust zones—for example, broad company context for an internal coding agent and a more restricted slice for another external client.

AiToolMap did not independently penetration-test those boundaries, so this remains a documented control rather than verified isolation performance.

Unabyss is attractive because it centralizes context. That is also its largest structural risk.

A well-used account may accumulate company strategy, project history, emails, meeting notes, CRM information, repositories, calendars, documents, personal preferences and identity information.

Even with read-only ingestion, this makes Unabyss a high-value concentration point. An attacker who obtains context access may not be able to modify Gmail, but they could learn far more than from compromising one ordinary productivity app.

This is why permission-aware retrieval, per-token redaction and account security matter so much.

The public documentation is encouraging, but AiToolMap found no fixed-panel independent penetration test, audit report or long-term incident history. The product is new enough that operational trust is still being established.

Teams should use the principle of least context: connect only sources that provide meaningful value, use separate credentials for different agents, and exclude personal/company-confidential material where the external client does not need it.

A portable memory layer is only useful if it can ingest the systems where context actually lives.

The current integrations catalog is already broad across engineering, communication, productivity, CRM, task management, meetings and advertising.³

The August changelog shows active development around multiple Google accounts, Drive sync scope, OAuth refresh reliability, onboarding and new MCP clients/integrations.¹ This matters because integration products often fail not at launch but after token expiry, provider API changes or permission drift.

The product’s public changelog specifically documents fixes for OAuth token refresh during long-running imports across several services and improvements to connection state.¹

AiToolMap did not independently test any connector, so integration reliability remains first-party evidence.

The MCP API reference gives builders concrete operational limits rather than hiding them behind vague “fair use.”²

The current docs list default per-credential limits such as 1,000/hour for normal read tools, 60/hour for `agentic_query`, 25/hour and 50/day per account for `store`, and 10/hour for certain export-creation operations.²

The API returns distinct errors for rate limiting, plan limits, memory unavailability and temporary service degradation, and can include retry hints.²

OAuth details are also specific: PKCE S256, access-token lifetime of four hours, refresh-token lifetime of 30 days, exact resource/audience matching and token-family revocation on replay outside a grace window.²

For a developer-facing infrastructure product, this level of documentation materially improves evaluability.

The current homepage publishes three main subscription tiers and a seven-day trial without a credit card.¹

Pro is displayed at $15/month on monthly billing or $13/month when billed annually. It allows up to three connected MCP agents, up to 20 connected accounts and a premium usage tier.¹

Max is $89/month or $79/month annually, with unlimited agents, unlimited connected accounts, multi-account support for Google apps, unlimited usage and early access to new features.¹

Team is $49/seat/month or $39/seat/month annually, with a minimum of three seats and team workspace/invoice/seat administration.¹

The Terms also retain a credit/pay-as-you-go framework for AI-powered operations and explain that credit consumption can vary with model, operation size/complexity and third-party provider costs.⁴ The current homepage, however, describes Pro/Max as subscription offerings with no additional costs in their listed usage framework.¹

Buyers should therefore rely on the current checkout/plan screen for the exact commercial model in force, particularly because the Terms support both credit-based and subscription-based mechanics.

For a solo user who genuinely moves among several agents, Pro’s current price is low compared with the time spent repeatedly rebuilding context. Max is a much larger jump and makes sense only for intensive AI usage and many connected accounts/agents.

AiToolMap audited all 50 members of `ai-review-panel-2026-08-v4` under methodology v1.8. No current exact-product Unabyss review, benchmark or qualifying verified-user population was located in the fixed panel.

That includes G2, Capterra, Gartner Peer Insights, TrustRadius, PeerSpot, SoftwareReviews, Trustpilot, app stores, mainstream technology publications and model/agent benchmark sources.

The absence of panel coverage does not mean Unabyss performs poorly. It means the product is still too new/niche to have accumulated a representative external evidence base in those sources.

Model benchmarks are also a poor fit. Unabyss’s value comes from ingestion, structuring, retrieval, permission boundaries and cross-tool distribution. A generic benchmark score for Claude, GPT or Gemini does not measure whether Unabyss retrieves the right context, avoids leaking excluded context or keeps source data current.

The first-party evidence, by contrast, is unusually rich. AiToolMap read current Terms and Conditions, the Privacy Policy, Security & Trust page, MCP API documentation, integrations catalog, homepage/pricing and status page.^4562317

Beyond the panel, Product Hunt confirms two 2026 launches and strong launch visibility, while Product Watch independently describes the same MCP-native context-layer concept.⁸⁹

Unabyss’s Terms and Security page both make a point that many context products obscure: once selected context is delivered to a third-party AI tool, the external tool’s own policies govern what happens next.⁴⁶

The Terms say distribution through MCP is initiated and controlled by the user, with granular control over what Context Data is shared. They also state that Unabyss is not responsible for how a third-party tool subsequently uses, stores, processes or displays the data.⁴

This is not merely legal boilerplate. It is the key architectural trust boundary.

A user can configure Unabyss carefully and still send context to an AI service whose retention/training/account terms are different. The context layer can control what leaves Unabyss; it cannot retroactively control a third-party processor after delivery.

For sensitive professional use, users should therefore create different MCP credentials/scopes for different tools rather than treating every AI client as equally trusted.

The current Privacy Policy is dated July 31, 2026 and identifies the controller as OneType Prosta Spółka Akcyjna in Warsaw, Poland.⁵

The policy covers account/contact data, billing, uploaded Context Data, OAuth authorization tokens, voice recordings, support correspondence, usage/device data and integrated-service data.⁵

For Google integrations, it explicitly lists read-only scopes and says Google user data is never used for advertising, never sold, and not used to train AI models.⁵

For AI processing, the policy identifies third-party AI systems including OpenAI, Anthropic/Claude and Google/Gemini. It says Context Data and conversation messages may be processed by those providers for AI output and semantic search, limited to requested functionality, and that Context Data is not used to train or improve external AI models.⁵

That disclosure is better than generic wording such as “trusted AI providers may process data.” It lets a procurement/security reviewer know which major model-provider families may receive context.

The policy also says ElevenLabs Scribe processes voice recordings for speech-to-text.⁵

Analytics/marketing technologies are separately listed: Google Analytics, Google Tag Manager, Meta Pixel, LinkedIn Insight Tag, X/Twitter Pixel and Microsoft Clarity.⁵ Users who are sensitive to marketing analytics should distinguish those website tracking flows from the private Context Data processing described for the application.

The product’s trust pitch is “your data, your rules,” and the legal/security materials provide meaningful support for that claim.⁶

The Security page says imported context remains until the user deletes it, even if an integration is disconnected; disconnecting stops future imports but does not silently erase previously built context.⁶

The Terms allow account deletion and say that deletion triggers an irreversible purge of Context Data and AI Output, subject to limited retention for identifiers, billing, tax, security and audit records where legally required or needed for claims.⁴

The Privacy Policy provides GDPR rights and identifies privacy contact information.⁵

The product also emphasizes exportability: markdown dossiers/Exports can be generated, edited, versioned and downloaded, and builder-facing material says users can export everything as Markdown rather than being locked into proprietary memory.⁴¹

That is important for a product whose job is to become a durable representation of the user. Portability reduces the risk that the context layer itself becomes a new vendor silo.

The current Security & Trust page states that data is encrypted in transit and at rest, access to production systems is restricted on a need-to-know basis, regular backups are maintained, raw file storage is placed behind a proxy layer, and source integrations use read-only scopes.⁶

The page says Unabyss’s frontend/landing surfaces run on Cloudflare’s edge network, while APIs run on hardened access-controlled servers.⁶

It also provides a security reporting address for vulnerabilities.⁶

Unabyss describes itself as GDPR aligned and says SOC 2 Type II is **in progress**, with controls/policies/monitoring implemented and the company engaged in the formal audit process.⁶

That wording should be preserved precisely. “SOC 2 Type II in progress” is not the same as holding a completed SOC 2 Type II report. Enterprise buyers should request the current audit status and available security documentation rather than treating an in-progress badge as certification.

A current status page provides at least some operational transparency. At the captured Aug. 25 update, `unabyss.com` showed 100% 30-day uptime and `app.unabyss.com` approximately 99.769%.⁷ This is a short vendor-operated window, not an independent SLA measurement, but it is preferable to no public service-health signal.

The Privacy Policy says OpenAI, Anthropic and Google/Gemini may process Context Data for requested AI functionality.⁵

Product Hunt’s “built with” surface says Claude Opus/Sonnet are primary models for Unabyss’s agentic-query reasoning, while ElevenLabs Scribe is used for speech-to-text.⁸

This is useful implementation context, but it still does not justify importing Claude benchmark scores into a Unabyss rating.

The quality bottleneck is often retrieval, not raw reasoning. If Unabyss retrieves the wrong source lines, the best foundation model will produce a confidently wrong context-aware answer. Conversely, precise retrieval can make a cheaper model more useful.

The product should eventually be evaluated on retrieval precision/recall, source freshness, permission leakage, context compression, citation/traceability and downstream task improvement—not on generic LLM intelligence leaderboards.

The current builder job posting even describes internal work on precise, traceable retrieval, permission-aware access controls and architecture evaluation across knowledge graphs/hybrid retrieval/structured extraction.¹ That reinforces that retrieval architecture is core product IP, but it is still first-party hiring material rather than benchmark evidence.

Product Hunt confirms meaningful launch interest. The original May launch won Product of the Day, #2 of the week and #5 of the month; the July “Unabyss for Claude” launch also ranked #1 of the day.⁸

Product Watch independently describes Unabyss as an MCP-native self-updating context layer and identifies the same basic value proposition.⁹

Those are useful adoption/discoverability signals, but neither is an independent reliability study.

AiToolMap did not locate a substantial verified-user review population in the fixed panel, nor a third-party test that measures whether Unabyss materially improves answer quality or reduces tokens/time relative to built-in memory or manually maintained context files.

A beyond-panel Stackbases review repeats several product claims and frames Unabyss positively, but it also relies heavily on vendor-described architecture/outcomes and does not provide a reproducible retrieval benchmark. AiToolMap therefore treats such directory/editorial evidence as context rather than a rating input.

The most important future evidence would be comparative testing: the same set of tasks with native model memory, manual context files, naive RAG and Unabyss, scored for factual correctness, context freshness, leakage and token cost.

Unabyss is most attractive for AI power users who actively use multiple vendors every day: for example, ChatGPT for research/writing, Claude for long-form analysis, Cursor/Codex for coding and another agent for operational work.

Founders are a natural fit because company context, decisions, investor/customer information and project history repeatedly need to appear across tools.

Developers benefit when coding agents need awareness of repositories, issue trackers, conventions and prior decisions without maintaining multiple static context files.

Agencies and GTM users may benefit from keeping client/pipeline/call context portable across assistants, but they should pay particular attention to segregation and external-client permissions.

Unabyss is a weaker fit for users who only use one AI product with sufficient built-in memory, organizations that prohibit centralized cross-system context stores, or enterprises that require a completed SOC 2 report before any sensitive-data connection.

A sensible rollout is incremental: start with non-sensitive/read-only sources, evaluate answer improvement, create different MCP credentials for different clients, then add more sensitive context only where the incremental benefit is clear.