AgentsKit-io/agentskit

AgentsKit is a large and well-planned toolkit for building AI agents with JavaScript and TypeScript. It includes a lightweight core, model adapters, an agent runtime, tools, memory, RAG, skills, evaluations, observability, sandboxing, MCP support, and interfaces for several frontend frameworks. Developers can use only the packages they need instead of adopting the entire system.

The repository has an impressive engineering foundation. It uses a monorepo with clear package boundaries, formal architecture decisions, independent package versions, and a published stability policy. The core package has reached version 1.0, while the rest of the ecosystem is labeled according to its maturity. The README even explains when developers should choose a simpler or more mature tool instead. That honest guidance is useful and helps set realistic expectations.

Testing and automation are major strengths. Current checks for CI, end-to-end tests, coverage, documentation, bundle size, CodeQL, dependency review, and Lighthouse are passing. Releases also include package checks, API snapshots, packed consumer tests, and supply chain controls. The security policy clearly identifies powerful features such as shell access, file tools, URL fetching, database queries, MCP connections, and code execution. It also explains that these tools need allowlists, sandboxes, supervision, and audit records.

Documentation receives unusual attention. The project has a full documentation site, live examples, an agent-friendly reference, package guides, contribution instructions, architecture records, issue templates, a public roadmap, and a clear MIT license. The quick start works without an API key, which gives new users a simple way to understand the runtime before connecting a paid model.

The main challenge is the size of the ecosystem. With more than twenty packages, several user interface frameworks, many integrations, and a large documentation site, new users may have trouble choosing a safe starting point. A small set of recommended paths would help. For example, the docs could offer one path for a basic chat app, one for an autonomous tool-using agent, and one for a production service with memory, evaluation, and monitoring.

The top-level README should also be checked against the current stability document. Some packages are shown as alpha in the README while the stability policy lists them as beta. Even if package badges are correct, this difference can confuse users deciding what is ready for production. The comparison table with other frameworks should also include dates, tested versions, and links to evidence because competing products change quickly.

Security guidance could go further by providing complete production examples. A hardened sample should show filesystem limits, network allowlists, secret handling, tool approval, rate limits, timeouts, memory isolation, logging, and sandbox cleanup working together. Prompt injection and tool output poisoning should have dedicated tests across RAG, MCP, memory, browser tools, and delegated agents. The public Ask Docs service should also receive a clearly documented security support level because it accepts remote requests and works with project documentation.

Overall, AgentsKit is one of the most complete JavaScript agent toolkits available at this stage of development. Its strongest qualities are modular design, serious automation, honest maturity labels, and excellent documentation. A simpler onboarding path, fully aligned stability labels, and more complete secure deployment examples would make the project easier to adopt with confidence.