Open source AI coding agent: what you take on

An open-source AI coding agent gives you three things a closed one cannot: you can read what it sends, you can point it at any model, and nobody can take it away or reprice it. For a lot of teams that settles the argument, and it should.

What you actually gain

1. Auditability. You can read the code that decides what leaves your machine. For anyone with a compliance obligation, this is not a nice-to-have - it is the difference between "we believe" and "we verified".
2. Model independence. Point it at a hosted frontier model today, a cheaper one next quarter, a local one when policy changes. The tool stops being a bet on one vendor's pricing.
3. Survivability. A funded startup can pivot, get acquired or shut down. A repository you have cloned cannot.
4. Fixability. When it does the wrong thing, you can look at why - and sometimes fix it - rather than filing a ticket into a queue.

What you take on

The bill you no longer pay a vendor, you pay in attention - and it is worth reading where that money actually goes before assuming self-hosting is cheaper. Someone has to hold the model keys and watch the spend. Someone has to tune the prompts and the tool permissions - the defaults in an open agent are a starting point, not a product. Someone has to decide when to take an update and what to do when it regresses.

On a team with an engineer who enjoys that, it is nearly free. On a team without one, it is a slow tax that nobody has named.

Who genuinely should not choose open source here?

A one-person team whose scarce resource is hours rather than money. The whole value of a managed tool is that somebody else absorbs the tuning, the model routing and the regressions - and if you are alone, those hours come directly out of the work you are actually trying to ship.

The inverse is also true: a team with a platform engineer and a compliance requirement should probably not be paying a vendor for something they can run and audit themselves.

The field, briefly

One thing to check before you commit

"Open source" covers several licences with different consequences, and some tools described that way are source-available rather than open under the OSI definition.1 If your reason for choosing open source is legal rather than aesthetic, read the actual licence rather than the badge on the README.

Did that land?

Three questions on the trade.

What does open source mainly change about the work involved?

It removes the tuning and operations work from the picture It moves that work from the vendor onto your side of the line It makes the underlying model cheaper to call per token It removes the need for a permission model around the tools

Who is the worst fit for a self-run open-source agent?

A team with a compliance obligation about where data goes A team with a platform engineer who enjoys running things A solo builder whose scarce resource is hours, not money A team that wants to swap models as prices keep moving

Why check the licence rather than the label?

Some tools called open source are source-available, which differs The licence determines which models the tool is permitted to call Most open-source licences forbid using the tool commercially The licence sets what the vendor may charge you later on