CAIN-42 today

All docs

The state of the technology as of 1 October 2026, in one page. Every statement here was checked against the live sites or the code on that date. When something is not live, this page says so.

CAIN stands for Cognitive Artificial Intelligence Network. It sits between an autonomous AI agent and the consequential things it does (tool calls, payments, record changes), decides on each action *before* it runs, has that decision certified by a quorum of independent nodes, and leaves a signed record you can verify without trusting us.

What you can use right now#

Surface	Where	Status
Decision API	POST https://cainstudio.online/fabric/decisions (API key)	Live
Demo, no account	POST https://cainstudio.online/fabric/try?scenario=safe-read	Live, 20 requests per hour per address
Python SDK cainstudio	pip install https://cainstudio.online/cainstudio-0.3.0-py3-none-any.whl	Live (hosted wheel; not on PyPI yet)
LangChain / LangGraph	cainstudio.langchain.protect([...])	Live (in the SDK)
TypeScript	copy the helper in TypeScript	No npm package yet
MCPGate (MCP proxy)	MCP, Self-hosted	Live hosted; self-hosted via Helm
Signed decision records	Decision record format, /verify.html	Live; verify offline
Free signup	/signup	Live (per-address rate limit; no captcha yet)
Book a demo	/book-demo	Live

How a decision is made#

Send the tool call before you run it. CAIN runs it through this chain, in order, and returns every stage's verdict. The stage names below are the ones in every response.

#	Stage	What it checks
1	identity	Who is asking: the API key resolves to a principal (a service or a named agent key).
2	intent	The intent the caller declared for this action (recorded; not enforcing on its own).
3	policy	Your policy (OPA) allows this path and payload.
4	authorization	The account's entitlement is active.
5	risk	Known attack patterns (prompt injection, exfiltration shapes) and the action's risk level.
6	trust	The agent's earned trust for this kind of action. Unknown trust never becomes ALLOW by itself.
7	verification	If you submitted a plan, its structure is checked.
8	actionproof	If you submitted a plan, ActionProof checks it before any step runs.
9	approval	Whether a human must approve; if so the call is held with an approval id.
10	consensus	The decision is ordered and certified by 3 of the 4 nodes of the PBFT cluster.
11	egress	Network destinations named in the action are inside your allowlist (observe-only until you add a rule).
12	toolargs	Arguments match the tool's registered schema (observe-only for tools without a schema).
13	artifacts	Pinned, publisher-signed tools, skills and MCP servers; dangerous capability combinations.
14	enforcement	Where the verdict lands (the hosted gateway, MCPGate or the SDK guard).
15	execution	Whether the call may execute: allowed, held, or prevented.
16	outcome	The action's lifecycle state.
17	evidence	The decision is recorded, signed, and bound to a hash of the exact arguments.

Stages 11 to 13 run after consensus and can only add a denial, never an allow. If one of them cannot run, it denies (fail-closed).

The rule for callers: run the tool only when verdict starts with ALLOWED and blocked is false. Everything else, including a timeout and a non-200 response, means "do not run". A new agent's first actions usually come back REQUIRE_APPROVAL.

Speed#

A hosted decision currently takes about 0.7 to 1.0 seconds, of which about 650 ms is the consensus round (measured 2026-10-01; see Benchmarks). An owner can put an agent on the snapshot fast path (POST /fabric/agent-keys/{id}/fast-path), which checks a quorum-committed authorization snapshot instead of waiting for a round. That path is enabled in production but has not yet been measured end to end for a customer agent.

The cluster#

The live consensus cluster cain-mr-02 has 4 nodes in Atlanta, Los Angeles, Miami and San Jose, all on one hosting provider, and survives the loss of any one node. It publishes signed agreement proofs every hour; see /live-cluster.html.

What is not live#

Be precise about these when you build on CAIN:

• Not on PyPI or npm. Use the hosted wheel above, or the HTTP API directly.
• CAIN is a decision point unless you route calls through it. An agent that

calls a tool without asking CAIN is not stopped by CAIN. MCPGate and the SDK guard are how the verdict becomes enforcement.

• Evolutions E9 to E42 are tested libraries, most of them not on the hosted

decision path. Their evidence bundles say so.

• /fabric/frontier/* was withdrawn on 2026-10-01 (unauthenticated writes to

a shared store) and returns 404.

• No third-party audit or certification yet, and no independent reproduction

of the evidence. See what we do not do yet.

• Demo requests are stored but not emailed until notification is configured.

Where to go next#

1. Quickstart: a first decision in five minutes. 2. Python SDK: @cainstudio.guard(), approvals, runs. 3. Use with your existing stack: LangChain, OpenAI Agents, CrewAI, MCP. 4. Decision record format: verify a decision offline. 5. 7-Moat architecture: the seven layers behind all of this. 6. Changelog: what changed and when.