Crabnet is a secure AI workspace for controlling what agents and users can see, use, and publish inside a business context.

How does Crabnet control MCP tools?

Each node or session can expose a limited set of MCP tools, such as search, current memory, publish, review, or list nodes.

Crabnet is for teams that want to give AI agents useful context without giving them broad access to company systems.

Give every AI the right context. Nothing more.

Connect repositories, documents and decisions once. Crabnet resolves permissions before retrieval and keeps evidence attached to every answer.

Request early access Watch 60-sec demo

Private beta · Free for pilot teams

Permissions resolve before retrieval. Evidence stays attached to every context package.

CAcme engineering

Task context

Checkout service

How do we roll back the checkout deployment?

3 sources allowed by node policy

Context ready

Freeze new deploys, restore the previous release tag, then verify checkout health before reopening traffic.

Evidence

rollback-runbook.mdrelease-policy.md

Fixed internal evaluation · July 2026

The same corpus runs after every retrieval change.

20/20

queries passed

100%

citation accuracy

permission leaks

See context governance in 60 seconds

Watch Crabnet connect sources, enforce policy, and produce a cited answer.

Open the 60-second demo

From scattered sources to governed answers

Connect sources

Bring repositories, docs and drives into one context graph.

Define access

Map roles and policies. Crabnet resolves them before retrieval.

Ask with evidence

Return answers with citations only from permitted sources.

Permission checks happen before retrieval

Blocked content is never retrieved, summarized or leaked into the model context.

Allowed context

payments-servicerollback.md

Product docsonboarding.md

Returned with citations

Blocked context

Financeforecast.xlsx

Peoplecompensation.pdf

Never sent to the model

Private beta — early access

Help shape safer AI context

Crabnet is a hands-on private beta for 5–10 engineering teams. It is free during the pilot, with direct onboarding and regular feedback sessions.

130-minute onboarding

2Use one real workflow

3Share candid feedback