Cortex
A self-hosted alternative to Dify and Flowise — design an AI agent's reasoning on a canvas, then watch it execute node by node.
Cortex is a visual canvas for designing how an AI agent thinks — not what it says next.
What the agent does is a diagram, not a prompt somebody edited last month.
Agent logic written as prose in a text box becomes unmaintainable faster than anyone expects. Six months in, nobody remembers why a particular instruction is there, whether removing it breaks something, or which of the four paragraphs is doing the work. Changes get made by adding another sentence rather than by editing, because editing is too risky. The agent slowly becomes a thing the business depends on and nobody is willing to touch, and the person who wrote it has usually moved on.
- Visual agent canvas (6-category nodes)
- Agent versioning (draft/published/archived + rollback)
Every published version is still there, and rolling back takes one action.
The worst day with an agent is the day someone improves it. A change goes out, quality drops in a way nobody can quite characterise, and the previous behaviour exists only in whatever the last person happened to remember. Teams respond by freezing: the agent stops improving because improving it is how it broke last time. What that costs is not the bad version, it is every good version that was never tried afterwards.
- Agent versioning (draft/published/archived + rollback)
- Conversation history (thread + node_trace)
The agent draws on your documents, and the trace shows which ones.
A generic assistant is confidently wrong about the specific things a business cares about — its pricing, its policies, its edge cases. Bolt on retrieval and a second problem appears: when the answer is wrong nobody can tell whether the document was missing, the wrong passage was found, or the model ignored what it found. Debugging becomes guesswork, and the usual outcome is that people stop trusting the agent and go back to searching manually.
- Knowledge base (upload/chunk/embed/vector search)
- Live-preview chat + node-execution trace
Nodes come in six categories covering what an agent needs to be told — the knowledge it may consult, the tools it may call, the logic that routes it, the constraints it operates under, and the point at which it escalates to a person.
The graph persists as structured data and round-trips cleanly, so an agent is a versionable object rather than a drawing.
Saves are version-checked rather than last-write-wins. Two people editing the same agent, or one person with the same agent open in two tabs, produce a conflict that is reported rather than a silent overwrite of whichever save landed first.
For a canvas — where a "document" is an hour of careful wiring — the silent-overwrite failure is the one that loses real work.
What streams back is not just the reply. Node execution streams alongside it — start, output and completion per node, plus errors — so the designer watches the agent reason rather than inferring it afterwards.
Each node carries a time limit and the run carries a total budget, so a misconfigured tool stalls one preview rather than hanging the session. Independent branches execute in parallel.
That is what makes a production complaint tractable: the report is "it gave a wrong answer on Tuesday", and the trace for that exact message says which knowledge chunk it retrieved and which branch it took.
Knowledge is scoped and metered per workspace: how many knowledge bases, and how many chunks in total.
That size discipline is the whole reason it is not React: a support widget that adds a third of a megabyte to a customer's marketing page is a widget their performance team removes.
Allowed domains are enforced on the server, not merely stored. A widget key lifted from one page does not work on another site.
Addresses are generated with a cryptographic random source rather than a predictable one, so a public agent's URL cannot be guessed by enumeration.
Metering an AI product after the fact is how a bill becomes a surprise. Checking before the response means the limit is a limit.
The platform underneath the product — counted in this product's own source, not claimed from a template. Where a row carries a list, open it to read the names behind the number.
Coinbase
dLocal
Flutterwave
Komoju
Mollie
NOWPayments
Opn
Paystack
ActiveCampaign
Customer.io
Segment
Canny
Amplitude
Custom Head Scripts
Pushover
Twilio SMS
fal.ai
OpenAI
Crisp
Tawk.to
Monday.com
Axiom
hCaptcha
Freshsales
Pipedrive
Salesforce
Pabbly Connect
AWS S3
Backblaze B2
Mux
QuickBooks
XeroHover a table to isolate what it touches; click one to read its relationships. The big nodes are what everything else hangs off. organizations is referenced by 46 tables because every single record in this product belongs to an organization, and that is what makes tenant isolation a property of the database rather than something the application has to remember on every query.
all 53 tables, as text
active_boosts · admin_notifications · agent_deployments · agent_templates · agent_versions · agents · announcement_dismissals · announcements · api_key_logs · api_keys · audit_logs · automation_log · automation_templates · conversations · credit_balances · credit_transactions · currencies · data_exports · data_imports · faq_categories · faq_items · feature_flags · inbound_webhooks · kb_chunks · kb_documents · knowledge_bases · languages · message_usage · messages · notification_log · notification_rules · notifications · organization_limit_overrides · organizations · plans · platform_secrets · platform_settings · product_purchases · products · provider_prices · push_devices · report_templates · roles · scheduled_tasks · subscriptions · theme_page_sections · theme_section_library · theme_sections · translations · user_sessions · users · webhook_logs · webhooks
The chassis, and anything particular to this product. 3 required, 2 optional.
Agent execution runs in Node runtime routes, which need database access. A single service — there is no separate execution worker to deploy.
Postgres with row-level security and vector search for knowledge retrieval.
Agent reasoning and the streaming preview run on it.
Documents are embedded on upload. Agents without a knowledge base do not need it.
Configured by the operator, not hardwired.
Already built — plug and play
Every provider integration in this list ships wired into the product. You bring your own keys, connect them in the admin, and go live — there is no integration code to write. Nothing is resold through us: the ongoing cost is whatever these providers charge you. Capabilities marked plan-gated elsewhere on this page unlock by plan, not by extra code.
What exactly do I receive after purchase?
The complete source repository behind the live demo — 113,964 lines across the application and its database, 53 tables defined by 50 migrations that the shipped scripts apply for you, plus the seed data, the demo accounts and the setup documentation. Not a subset and not a scaffold: the same code that runs the demo you just used.
Is the live demo the same product whose source I receive?
Yes. The Cortex demo runs the delivered release (v1.0.0), and the build figures on this page are measured from that same source — not from a showcase build kept separately.
Can I test every account role before buying?
Yes — all 2 of them: Workspace user, Platform admin. Each one is seeded in the live demo and reachable from the role board above, so you can inspect the product from the customer's side, the agent's side and the administrator's side before you decide.
Is this a complete product or a starter template?
Complete. 76 page modules and 233 HTTP operations, with authentication, roles, billing, an administration panel, background jobs and seeded demo data all wired and running. A template gives you the shape of an application; this is one you can deploy and start operating.
Can I rebrand and modify it?
Yes, without asking. The name, identity, copy, styling and code are yours to change. Rebranding and operating it as your own service is exactly what the licence is for.
What does the licence allow?
A perpetual, worldwide, non-exclusive, non-transferable licence to the product you bought and to every update released for it. The Standard licence covers one business — yours. If you deploy for clients, the Client licence covers up to five. You can modify it, rebrand it, deploy it and charge your own customers.
Can I resell the source code?
No. Reselling, redistributing, sublicensing or giving away the source itself is not permitted under either licence. You build and deliver products with it — the code stays with you.
What services and ongoing costs are required?
You bring your own accounts, so we never mark anything up and there is no ongoing cost to us beyond the purchase. Required: Node host, Supabase, An AI provider key, An embeddings provider, A payment provider. Optional: none. The table above says exactly what stops working without each one; most have usable free tiers at low volume.
How difficult is deployment?
Two ways, and most buyers use the first. The installer in your saascode dashboard walks it: you create free accounts at Vercel and Supabase, connect them, and it provisions the database, applies 50 migrations, seeds the data, deploys and verifies the result — you watch the steps go green. Or take the source and do it yourself with the shipped scripts and docs, which is the same sequence run by hand.
What updates and support are included?
Updates to that product for as long as we maintain it — no renewal fee and no expiry date on your access. Support covers download problems, defects in the code as shipped, questions about what the product does and how it is structured, and licence questions. It does not cover debugging your own modifications, building features for you, or setting up your hosting and third-party accounts.
Why does a complete codebase cost $179?
Because it was built once and is distributed many times. SaaSCode builds each product and ships the same finished release through the catalog, so the price reflects repeatable distribution — not a reduced codebase, an unfinished template, or what the same work would cost commissioned.
Do I need to be a developer?
No, for the normal path. You create free accounts at Vercel and Supabase, connect them to the installer in your saascode dashboard, and it does the deployment for you — provision, migrate, seed, deploy, verify. What it asks of you is creating two accounts and copying a token, not writing code. If you would rather deploy it yourself, or host it somewhere else, the full source and the scripts are yours to do that with. Running the product day to day needs no technical skill at all: that happens through its own interface, which is what the demo shows.