Takibi Base launches cited knowledge base for AI agents

Takibi Base has launched a shared knowledge base that returns exact cited passages to AI agents through CLI or API instead of generating answers. The product is aimed at teams that want tighter access control, auditable evidence, and less context loss across agent workflows.

Why it matters: - AI agents often lose track of useful context across documents, notes, and sessions. - Takibi Base is designed to keep source material in one library and return evidence that can be checked against the original documents. - The system is built to reduce overexposure by limiting each agent to only the context its job needs.

What happened: - Takibi Base launched its shared knowledge base on Sept. 29, 2026. - The product lets AI agents retrieve exact cited passages from company documents through the Takibi CLI or authenticated API. - The platform does not generate prose answers. - Deian Isac, founder of Takibi Base, said the product keeps source material in one library and returns evidence with citations instead of text an agent made up. - Takibi Base is available at takibibase.com and through the app at app.takibibase.com.

The details: - One human operator curates documents in projects and folders. - Profiles control what each agent can read. - Each Profile holds the permissions for its keys, so separate tools can see separate views of the same workspace. - Keys can be rotated or revoked. - Pausing a Profile stops all of its keys at once. - When an agent asks a question, Takibi returns matching passages with citations and an evidence support score. - When no matching evidence exists, Takibi returns no passages. - Notes are untrusted, Profile-scoped scratch with a fixed 30-day lifetime. - A person reviews notes in a queue before anything useful becomes a draft sourced from real documents. - Notes never become trusted sources on their own, are never shared across Profiles, and are never promoted automatically. - Conversations let agents exchange context around shared work. - Tasks give work an owner, let agents claim work, track progress, attach result links, and submit for review before the workspace owner accepts the work as done. - Projects and their structure can be exported in Open Knowledge Format. - Original files can be downloaded at any time.

Between the lines: - Takibi Base is betting that agent systems need traceable evidence more than fluent answers. - The access model suggests the company is targeting teams that want shared knowledge without giving every tool the same permissions. - The separation between trusted sources and scratch notes points to a workflow built for auditability, not just convenience. - The product also reflects a broader push to make agent output easier to review and verify in business settings.

What's next: - Takibi Base said it has shipped several recent updates since Sept. 19. - Those additions include document uploads from the CLI, per-Profile task and note permissions, editable usage budgets, a notes review queue with source drafts, a one-command install that teaches an agent to use Takibi, and an in-app helper named Tibi. - The company is offering two plans with a seven-day trial and card required at checkout. - Ember costs 18 euros per month for two Profiles, 250 asks per day and 300,000 tokens per day. - Campfire costs 52 euros per month for eight Profiles, 800 asks per day and 800,000 tokens per day. - Prices exclude VAT. - Billing starts after the trial unless cancelled.

The bottom line: - Takibi Base is trying to make AI agent work auditable, permissioned and tied to exact source evidence instead of generated guesswork.