AgentDock

The trait that decides it early is that you see something running. A complete reference client comes with the library, so the first hour is spent using an agent rather than assembling one, and for a TypeScript team evaluating options that shortens the comparison considerably.

What you should expect is rough edges, because the project describes itself as early and the client is a reference rather than a product. Pick it if you enjoy being on the front of a codebase. Pick VoltAgent when you need the thing you build this month to still compile next month.

The dependency story is the problem. Without a released artefact there is no semantic version to constrain, no changelog boundary and no way to distinguish a breaking change from a Tuesday, so every upgrade is a diff review. The documented install is a clone and a package-manager run against the working tree.

What it does right is concurrency. Sessions are isolated from one another, so two runs in the same process cannot contaminate each other's state, and that is a property most young frameworks discover only after an incident.

1. Making tool availability a function of conversational phase is the correct response to two known problems at once: schema bloat in the prompt, and models reaching for a capability that is inappropriate at that moment. Most frameworks pass everything on every turn and hope.

2. An evaluation framework ships with the library, which is unusual and welcome. 3. No results from it are published, so whether the phased tool exposure improves outcomes over a static set remains an open question the project is equipped to answer and has not.

A permissive core plus a promised cloud is a common structure and this one is at the riskiest point of it: the giveaway is shipping and the revenue is not, so there is nothing yet testing whether anyone will pay. 1,713 stars measures curiosity about the free half only.

Moat: none established. Likely path: either the hosted product arrives and defines the business, or the project stalls at the point where maintaining a framework stops being interesting without income. Position: watch it, prototype on it, and do not assume the cloud arrives on the schedule you would like.

Cost is inference and staff time, which makes the budget conversation short. Everything after that is absent. There is no scheduled or background execution mode, so whatever is built here runs inside an application my team writes, deploys and carries the pager for.

The storage abstraction is pluggable, which at least means our own key-value and vector estate can back it rather than a vendor's. Onboarding for a TypeScript engineer is a few days. Not yet: I will not put an unreleased core underneath a service my team has to support.

The extension model is the right shape. Everything is a node, tools are nodes, models come from a registry I can add to, and the storage layer is an interface rather than a hardcoded database. A permissive licence over readable TypeScript means I can change any of that without asking.

What is missing is the protocol. The integration is documented as planned, which is the polite version of absent, so every tool I already run needs a wrapper written here. That is the work I most resent doing twice.