KaibanJS

The daily trait is visual. Instead of reading a scrolling log you watch each task move across a board in real time, and for a JavaScript team that has never wanted a Python sidecar in the stack, that alone justifies an afternoon of trying it. Roles and goals are declared in a few lines and the first crew runs quickly.

Where it stops is depth. There is no unattended mode, so it is a thing you sit and watch. Pick it for demos, internal tools and teaching. Pick VoltAgent when the same agents have to run without anyone looking at the board.

The architectural gap is protocol. The framework supports neither side of MCP, which means every tool an agent reaches for is integration code somebody on your team writes, tests and maintains. Competing frameworks inherit a growing catalogue of servers for free. This one inherits nothing, and that difference compounds every quarter the ecosystem grows.

The consequence is a widening maintenance tax on the least interesting part of the system. The one thing done right: agent and workflow state live in a single Redux-inspired store, so at any moment there is one place to look when a run goes sideways.

1. An agent is a role plus a goal plus a tool list. 2. A task declares its expected output, which is the closest thing here to a verification contract, since the expectation is stated before the model runs. 3. Completed results are piped into later tasks, making the dependency graph explicit rather than emergent. 4. The visible board is a projection of one state store, not a separate system.

No benchmark is published and the documentation is thin on how expected outputs are actually checked. Declaring the expectation is not the same as testing it.

Adoption is the number that matters and it is small: roughly 1,470 stars since a July 2024 debut, against JavaScript agent frameworks with an order of magnitude more attention. Nothing is sold, nothing is hosted, and no funding signal appears anywhere on the project. That is a side project with good taste, not a venture asset.

Moat: none. The abstraction is copyable in a weekend by anyone who wants it. Likely outcome is not acquisition but attrition, or absorption of the board idea into a larger framework's observability layer. Position: pass on the company, borrow the interface.

The board demo lands well in a room. Procurement then finds nothing to buy, which is the good news, and nothing to rely on, which is not. There is no unattended execution mode, so this can never sit in our pipeline and produce an artifact overnight. It stays a thing an engineer runs on a laptop while watching.

No single sign-on, no audit trail, no retention policy, and support is a small repository with one maintainer group behind it. Onboarding is genuinely cheap for a TypeScript team, perhaps half a day. Not yet. Revisit when it can run without a human watching a board.

MIT and small enough to read in an evening, which is the licence and the size I want from anything I might have to patch myself. Install is npm install kaibanjs, or the init scaffold if I want the example project. Keys are mine across the three hosted vendors it speaks to, so nobody is reselling me inference.

The part that annoys me is the model layer. There is no local endpoint setting, so the machine I built specifically for this cannot participate. In a JavaScript framework that is a fifty-line pull request. I would rather fork it than wait, and the licence lets me.