Packages overview

Eight packages publish to npm under @tailored-ai/*.

Package	One-line	Install when
`@tailored-ai/cli`	The `tai` command. REPL, one-shot, project mode, bundled web UI.	You want to run TAI as a user.
`@tailored-ai/core`	Agent runtime, tools, providers, channels, persistence. Framework-free TypeScript library.	You're embedding the agent in your own Node service.
`@tailored-ai/server`	HTTP API: Hono routes, SSE streaming, webhook intake, optional static-UI mount.	You want the HTTP surface without the CLI's REPL.
`@tailored-ai/browser-mediator`	Framework-agnostic bounded browser tool. Egress allow-list, vault refs, opaque ids, HITL gates.	You want a safe browser tool for any agent (TAI, OpenAI function-calling, Anthropic tool-use).
`@tailored-ai/trusted-actions`	HITL approval gateway for risky actions. Mobile PWA, web-push, sandboxed Docker executor.	You want an agent to do things (place orders, submit forms) with human sign-off.
`@tailored-ai/provider-bedrock`	AWS Bedrock model provider. Converse API, tool calling, AWS credential chain.	You want agents on Bedrock-hosted models (Claude, Nova, Llama, Mistral).
`@tailored-ai/provider-openrouter`	OpenRouter model provider. One key, hundreds of models, streaming + model discovery.	You want to switch between hosted models without juggling vendor keys.
`@tailored-ai/provider-anthropic`	Anthropic Messages API provider. Prompt caching, beta headers, streaming.	You want Claude with caching + the knobs the minimal built-in lacks.
`@tailored-ai/provider-openai`	OpenAI provider. Reasoning-model handling, org/project headers, streaming.	You want GPT/o-series models without the built-in's parameter pitfalls.
`@tailored-ai/channel-slack`	Slack channel plugin (Bolt, Socket Mode).	You want to talk to your agent from Slack.
`@tailored-ai/google-tools`	Gmail, Google Calendar, Google Drive tools.	You want agents reading and acting on your Google account.

Two more packages live in the monorepo but stay private:

@tailored-ai/ui: the React + Vite SPA. Built and bundled inside @tailored-ai/cli. Not published standalone.
@tailored-ai/site: this documentation site. Built and deployed to GitHub Pages from main.

Dependency graph

@tailored-ai/cli             depends on  @tailored-ai/core
                             depends on  @tailored-ai/server
                             bundles     @tailored-ai/ui (built into ui-dist/)

@tailored-ai/server          depends on  @tailored-ai/core

@tailored-ai/core            depends on  @tailored-ai/browser-mediator
                             (for the browser_mediator tool)

@tailored-ai/trusted-actions depends on  @tailored-ai/core
                             (for the request-action tool wiring)

Read that as: install @tailored-ai/cli and you transitively get everything except trusted-actions. Trusted-actions is intentionally separate because it runs in its own Docker container; installing it alongside the agent is opt-in.

Pick the smallest thing that works

The temptation is to install the catch-all. Don't.

End user running an agent: @tailored-ai/cli only.
Embedding into your own backend: @tailored-ai/core, plus @tailored-ai/server if you want the HTTP routes.
Just the browser tool, for a non-TAI agent (Claude, GPT, your own framework): @tailored-ai/browser-mediator only. Zero dependency on the rest of TAI.
Just the trusted-actions executor, even without TAI: @tailored-ai/trusted-actions runs standalone.

Versioning

All five public packages release together. They share one version line because they evolve together; a breaking change in @tailored-ai/core rolls all of them. See RELEASING.md in the repo if you're publishing a fork or your own scope.

Optional peer dependencies

A few tools in @tailored-ai/core are gated by optional peers. They're listed in Installation → Optional dependencies. The short version: playwright, md-to-pdf, pdf-parse, and tesseract.js aren't installed by default. Add them only when you enable the corresponding tool.