Open Source Manufacturing ERP/MES/QMS: Carbon Reviewed

Carbon puts ERP, MES and QMS on one Postgres schema and lets you run it yourself. Here is what that covers, what is gated, and how to test the claims.

Binary Buddies · 5 min read · 12 views

Manufacturing software usually arrives as three products. The ERP holds BOMs and costs, the MES runs the shop floor, and the QMS holds NCRs and calibration records. Each has its own database, and the seams between them are sync jobs. The seams are where traceability breaks. When a customer asks which material cert went into serial 0412, someone ends up joining exports from three systems by hand.

The second problem is where the data lives. If your production records include controlled unclassified information, 'the vendor's cloud' may not be an acceptable answer. Open source helps here only if you can run it yourself, read the code, and keep the database inside your own network.

Carbon is an open source ERP, MRP, MES and QMS that makes that claim. This article goes through what its self-hosting page actually commits to, what is gated behind a commercial license, and what to test before moving a shop onto it. It is based on the vendor's own documentation, not a deployment we ran, so treat the claims as things to verify.

What is open and what is not

The Community edition is AGPL-3.0, and the source is on GitHub. The page says you can inspect the application, run your own security review, and extend it for your process. AGPL is a real constraint if you plan to modify it and expose it to users outside your organization. Have someone read the license against your plans before you fork.

Several capabilities sit outside that edition. The page states that air-gapped deployment and licensing are Enterprise features. It also states that self-hosted API keys and MCP require a commercial license. Enterprise BYOC engagements can add architecture work, installation, legacy data migration, training and an SLA.

So 'open source' here means the core application is readable and runnable. It does not mean every deployment mode is free. If your requirement is an isolated network with no outbound calls, you are in the commercial tier from the start.

Getting a stack running

The documented quick path is short:

git clone https://github.com/crbnos/carbon.git
cd carbon
docker compose up -d
# App, API, MCP server and Postgres on your box.

That brings up the application, the API, an MCP server and Postgres. It is a sensible way to evaluate: load a real BOM, build a traveler, and see how the data model handles your parts before anyone discusses pricing. The page describes Docker as a starting point on one host, with a move to a clustered deployment later.

The page does not give hardware sizing, a backup procedure, or an upgrade path for the Compose setup. Those are in the linked deployment guides, which the page refers to but does not reproduce. Read them before you decide a single Docker host is enough for production.

One Postgres schema instead of sync jobs

The architectural claim that matters most is that ERP, MRP, MES and QMS share one Postgres schema with row-level security. The page says there are no synchronization jobs between separate product databases. The multi-entity section adds the operational consequence: one schema, one backup, one system to secure.

That removes a whole class of failure, such as a work order that closed in the MES but never reached the ledger. It also moves the risk. Row-level security policies become your permission boundary, so read them as part of any security review. A single database is also a single thing to restore, so test a restore before go-live, not after the first incident.

Files are handled separately. Attachments, drawings and certificates live in object storage you control, behind signed URLs and application access rules. That means your storage configuration is part of the security perimeter too.

Traceability as a data model

The quality pitch is that you open a serial number and see its material certificates, operators, measurements and deviations in one place. First articles, NCRs, CAPAs and calibrations link to the same production records. The page lists serial and lot genealogy in both directions, an NCR-to-CAPA workflow with sign-off, and certificates generated from live data.

This is the part to test hardest, because it is where a unified schema should pay off. Build a small assembly with a purchased lot, a machined part and a deviation. Then answer the recall question both ways: which units used this lot, and which lots went into this unit. If both answers come back from one screen without a manual export, the shared schema is doing its job.

Running the shop floor on your own servers

The execution layer covers digital travelers with work instructions, operator terminals, QR and barcode tracking on every unit, and finite-capacity scheduling. The page makes a specific availability claim for self-hosted installs: production does not depend on Carbon's managed cloud.

That claim is testable. Run a traveler on a terminal, then cut the connection between your site and the internet and see what still works. The answer will depend on your deployment mode and on anything the instance calls out to, such as an identity provider or an AI model endpoint. The page says you set the identity provider yourself, so an externally hosted IdP is an outage dependency you introduce.

Four ways to deploy it

The page describes one codebase across four models:

  • Managed BYOC: Carbon deploys into your cloud account and operates upgrades, certificates and monitoring. Application data stays in your account.
  • Docker: application, API and MCP server in containers with Postgres, on one host or clustered.
  • Self-managed cloud: your AWS, Google Cloud or Azure VPC with managed Postgres. You own the network, keys, backups and releases.
  • On-prem and air-gapped: no outbound calls inside an isolated network, an Enterprise feature.

The trade-off is operational. Each step toward self-managed gives you more control and hands you more of the work: certificates, monitoring, upgrades, key rotation. BYOC splits that burden, with the vendor operating releases while you keep the account and network. The page also says BYOC deployments are single-tenant, with their own application, database and object storage, and that you can ship a customized fork through the same staging and production pipeline as the standard build.

Multi-site accounting and the CMMC question

For multi-site manufacturers, the page claims multiple entities and locations on one ledger. Currency, chart of accounts and tax are configured per entity, with intercompany transactions, consolidated reporting and inter-site transfers. Check that against your actual structure, particularly if your intercompany flows or tax rules are unusual.

The compliance angle is aimed at CMMC and NIST 800-171 work. Keeping CUI in infrastructure you control is the stated reason to self-host. Enterprise BYOC deployments include system-specific inputs for your SSP, POA&M and SPRS score. Those are inputs to your documentation, not a certification. Running software inside your boundary does not make your organization compliant, and the page does not claim it does.

Bringing your own models

The AI position is that you use the REST API and the built-in MCP server with hosted or local models. Requests stay inside your deployment and are subject to Carbon permissions, and you choose the models. For a shop with data-handling rules, a local model behind the same permission layer is the safer shape.

There is a tension in the page's own text. The Docker description includes the MCP server, yet self-hosted API keys and MCP require a commercial license. Expect to run the server in the Community stack but need a license to use it self-hosted. Confirm that with the vendor before you build any agent workflow on it.

Where this does not fit

If you need air-gapped operation, self-hosted API keys or MCP, or vendor-led migration, you need a commercial agreement, so the cost comparison is not free software against a quote. If your team has nobody to run Postgres, object storage, TLS and upgrades, a self-managed deployment moves that work onto you, and BYOC may be the realistic option.

Be careful with the AGPL if you plan heavy customization, and read the row-level security policies instead of trusting the phrase. The page gives no sizing figures, no benchmark numbers, and no migration tooling details for legacy systems beyond offering the service. Anything on those points has to come from the documentation or a proof of concept on your own BOMs, your own travelers, and a real restore test.

  • #Carbon
  • #MES
  • #Open Source
  • #QMS
  • #manufacturing ERP
  • #self-hosted ERP

Comments

Add a comment
Sign in with Google to join the conversation