01
Token separation
Durable OAuth tokens stay in your credential layer. Agents never hold them; CybrLink keeps them out of the model runtime entirely.
OAuth and app access for SI agentsGive agents app access.
Give agents app access.
Keep tokens out of the model.
CybrLink is the control layer between your SI (super intelligence) agents and user-connected apps. Run it with a credential layer you control so OAuth tokens stay in your infrastructure while agents get scoped, auditable access to 702 catalog integrations.
55 typed MCP tools cover 11 of those apps. The rest of the catalog connects through the scoped proxy. We publish a tool count only where a schema exists.
- Catalog integrations
- 702
- Typed MCP tools
- 55
- Tokens in the model
- 0
- Calls audited
- 100%
Watch CybrLink in action
One prompt. Real app access. Zero tokens in the model.
CYBRLINK_SEARCH_TOOLS
⌕ read unread email, post summary to slack3 found
GMAIL_LIST_MESSAGES
List unread messages in the inbox
GMAIL_GET_MESSAGE
Read one message by id
SLACK_POST_MESSAGE
Post a message to a channel
PLAN
1 List unread email threads
2 Summarize with the agent's model
3 Post summary to #digest
GUARDRAILS
! Key scope: Gmail read-only
! Write access: Slack #digest only
AGENT_SESSION
claude code
claude codeSummarize my unread email and post it to #digest in Slack.
CYBRLINK SEARCH TOOLS ›
CYBRLINK SCOPE CHECK ›
CYBRLINK PROXY CALL ›
Read 14 unread threads and posted the summary to #digest.
Done — and the OAuth tokens never left your vault.
Reply…▍claude-opus-4-8 ↵
CYBRLINK_CONNECTIONSusr_9x2kLm7
Google Workspace
OAuth 2.0
Slack
OAuth 2.0
CYBRLINK_PROXY_CALLses_7k2m
SLACK_POST_MESSAGE
channel#digest
threads read14
✓ 200 OK · token stayed in vault
AGENT_CONFIG
AGENTClaude Code
MODELclaude-opus-4-8
TOKENS HELD BY AGENT0
Works withClaude Code
Codex
Grok Build
OpenCode
Hermes
OpenClaw
Paperclip
Claude CodeCodexHermesOpenClawPaperclipThe catalog
702 apps you can plug your agents into.
OAuth and API-key apps — connect once, and agents call them through one scoped key. Credentials never touch the model.
55 typed MCP tools across 11 apps. Catalog size and typed-tool coverage are different numbers.
A few lines of Python. Your agent is connected.
Install the SDK, mint a session, and route agent tool calls through a credential layer you control. Every one of the 702 catalog integrations uses the same call shape — and the provider token never reaches the model. Typed MCP tools (55 across 11 apps) are a smaller set than the catalog.
$pip install cybrlink
- One SDK call to mint a scoped session — no OAuth plumbing.
- Tokens never enter the model; the agent only holds a short-lived session.
- Same call shape for all 702 catalog integrations — no per-provider glue code.
- 55 typed MCP tools across 11 apps. Other catalog apps use the scoped proxy until a schema is published.
# pip install cybrlink
import os
from cybrlink import CybrLink
link = CybrLink(api_key=os.environ["CYBRLINK_API_KEY"])
# 1. Mint a scoped, short-lived session for one of your end users.
session = link.sessions.create(user_id="u_42")
# 2. Your agent calls a connected app through CybrLink. The user's
# OAuth token stays in your infrastructure — the model only ever
# sees this scoped session, never the credential itself.
inbox = link.proxy(
provider="google-workspace",
session_token=session.token,
method="GET",
endpoint="/gmail/v1/users/me/messages",
)
for message in inbox["messages"]:
print(message["id"])You bring the credential layer. CybrLink makes it safe to hand to an agent.
Self-hosting a vault (we support Nango today) gets you token storage. It does not get you a safe way to point an autonomous agent at it — that's the layer CybrLink adds.
02
Policy & per-user scoping
Every agent call is checked against the user, integration, and permission set it's allowed to use.
03
Agent gateway
One controlled API for tool access, connection state, and revocation, so agents never talk to the vault directly.
04
Audit trail
A reviewable record of which user, key, integration, endpoint, method, outcome, and failure reason each agent action produced.
Security controls
Controls that keep provider credentials out of the agent.
These are product controls. CybrLink does not claim SOC 2, ISO 27001, or GDPR certification. No audit report is published in this repository.
Credential vault
Provider tokens stay in the credential host. Agents call CybrLink and do not receive raw OAuth tokens or API keys in tool results.
Scoped keys
Each cbl_ key is limited to providers you grant. Writes are off unless you opt in, and short-lived cks_ agent sessions are read-only.
Audit trail
Proxied calls are logged as metadata: user, key, provider, endpoint, and status. Request and response bodies are not written to the log.
- API keys and cks_ sessions are stored as HMAC hashes and shown once
- TLS in transit; provider tokens are not stored in this app
- Append-only activity log with a per-row integrity HMAC
- Per-user connection checks before any proxied call
6 agent workflows with scoped app access.
Each recipe shows the apps, scopes, and starter prompt an agent needs — without asking for broader access than the workflow requires.
Email Agent
Triage the inbox, draft replies, ship follow-ups. Reads & writes across mailboxes with scoped auth.
Calendar Agent
Books, reschedules, holds prep blocks. Sees free/busy across team calendars with read-only fallback.
Code Agent
Watches deploys, incidents, and database health. Turns production signals into scoped operational action.
Build with custody from day oneConnect every app your agent needs.
Connect every app your agent needs.
Keep credentials out of the prompt.
Start with the catalog and connect flow. Pair CybrLink with your controlled credential layer when you are ready to keep token storage under your own control.