Which SKU do I need?
One screen to pick from the 14 published `@pleach/*` packages. Keyed by what you're building, not by SKU.
The @pleach/* surface is 14 published packages (plus
@pleach/observe in alpha and the reserved @pleach/trust-pack).
Most projects need three or four. The packages are branches grafted onto one
lattice — pick the branches that carry weight for what you're
building, leave the rest for later. This page maps what you're
building to the shortest install list, plus a one-line reason
for each. If you're not sure, start with the first row —
@pleach/core alone covers more cases than you'd guess.
The shortest answer
| You want | Install | Read |
|---|---|---|
| A streaming chat handler in 5 lines | @pleach/core | Getting started |
| A chatbot with a working UI | @pleach/core, @pleach/react | recipes/simple-chatbot |
| RAG over your knowledge base | @pleach/core, @pleach/react (+ your vector store) | recipes/rag-chatbot |
| An agent with tools | @pleach/core, @pleach/base-tools, @pleach/tools | Tools, @pleach/base-tools |
| A coding agent (SWE-bench-shaped) | @pleach/core, @pleach/coding-agent, @pleach/sandbox | @pleach/coding-agent |
| A multi-tenant SaaS with per-tenant cost rollup | @pleach/core, @pleach/gateway | Multi-tenant, Gateway |
| OTel-instrumented observability | @pleach/core (OTel is built in); @pleach/observe (alpha) adds destination plugs | @pleach/observe |
| PII / PHI scrubbing for HIPAA / GDPR / PCI-DSS | @pleach/core, @pleach/compliance, @pleach/compliance-contract | Compliance, Scrubbers |
| Regression-grade evals against the audit log | @pleach/core, @pleach/eval, @pleach/replay | @pleach/eval, @pleach/replay |
| An MCP server backed by Pleach sessions | @pleach/core, @pleach/mcp | @pleach/mcp |
| A LangChain / LangGraph adapter layer | @pleach/core, @pleach/langchain | @pleach/langchain, Migrating from LangChain |
The recipes column on the right is the canonical worked example
for each row. Most of them are under 100 lines of glue over
@pleach/core.
Shortcuts by shape
If "what are you building?" is too narrow a question, here are
seven common shapes the runtime is built around. Each maps to a
recipe in @pleach/recipes and a SKU bundle.
SaaS adding a chat tab
A typical multi-tenant web app that wants a chatbot tab. You own your auth, your DB, your billing. Add Pleach without touching any of them.
Install: @pleach/core, @pleach/react
Recipe: simpleChatbot
Reach for next: @pleach/observe (alpha) once you want
per-tenant cost rollup.
Vertical agent (specific tools, specific knowledge)
You're shipping an agent for a specific domain — legal review, medical coding, ops triage — with a curated tool set and a knowledge base. Not generic ChatGPT-shaped.
Install: @pleach/core, @pleach/base-tools, @pleach/tools
Recipe: verticalAgent
Compliance-bound (regulated industry)
HIPAA, GDPR, PCI-DSS, SOC 2. You need scrubbers on inbound content, audit attestation on every call, and a hash chain you can prove to a regulator.
Install: @pleach/core, @pleach/compliance,
@pleach/compliance-contract
Recipe: compliantChatbot
Reach for next: @pleach/replay once you need
chain-verifier attestation.
On an existing Anthropic / OpenAI Enterprise contract
You already have an Anthropic Enterprise or OpenAI Enterprise contract — SSO, ZDR, Workspaces / Projects. You want per-axis cost rollup inside one Workspace without a new vendor.
Install: @pleach/core, @pleach/gateway (+ @pleach/observe, alpha)
Recipe: enterpriseAgent
Read first: Migrating from Anthropic Enterprise
or Migrating from OpenAI Enterprise.
Multi-tenant platform (per-customer BYOK + billing)
You're running multiple customer agents on the same substrate, with per-customer BYOK, per-customer policies, and per-customer billing.
Install: @pleach/core, @pleach/gateway,
@pleach/compliance
Recipe: enterpriseAgent
Subagent swarm (bounded fan-out under a root turn)
A root agent that spawns N specialist sub-agents and reconciles their answers, with a per-root-turn cost ceiling. Common shape for research, planning, and refactor agents.
Install: @pleach/core, @pleach/recipes
Recipe: subagentSwarm
Regression-grade eval lab
You want fixture-driven eval against the audit log with replay determinism and Welch t-test scoring. CI-runnable, not notebook-shaped.
Install: @pleach/core, @pleach/eval, @pleach/replay
Recipe: evalLab
When you don't need a sibling SKU
@pleach/core already ships these in the substrate; you don't
need a sibling for them:
- Session lifecycle, channels, checkpointing. Built-in.
- Storage adapters (Memory / IndexedDB / Supabase). Built-in.
- Family-lock, model resolution matrix, transport selection. Built-in.
AuditableCallrow + event log + hash chain (schema-side). Built-in. The verifier ships in@pleach/replay.- OpenTelemetry spans (session.turn, llm.invocation, etc.). Built-in.
@pleach/observe(alpha, not yet on npm) adds destination plugs (Postgres, Supabase, Pleach Hosted); core ships the OTel emission. - Plugin contract. Built-in. Authoring conventions in Plugin authoring standards.
- The four-stage lattice + singleton synthesize seam. Built-in invariants enforced by lint on the core repo.
Quick "do I install this?" gut-checks
A few SKUs get asked about more than others. Short answers:
@pleach/base-tools— install if you want a working agent today and you don't have a domain tool set yet. Ships math, datetime, scratchpad, unit_convert, text_search, opt-in url_fetch.@pleach/tools— install if you're authoring tool definitions. Provides the contract + Zod helpers + queryable registry. Skip if you're using@pleach/base-toolsdirectly.@pleach/sandbox— install if you're running untrusted code (coding agent, user-supplied scripts). Contract today; vendor adapters land under@pleach/sandbox-<vendor>.@pleach/recipes— install if you want high-level factories (simpleChatbot,ragChatbot,compliantChatbot,verticalAgent,subagentSwarm,enterpriseAgent) instead of wiringSessionRuntimefrom scratch.@pleach/trust-pack— reserved name. Not shipping code today. Don't install.
Where to go next
Packages
The canonical per-SKU status table — versions, links, contract status.
Recipes
End-to-end runnable patterns — Next.js chat, OTel, multi-tenant, compliance.
Comparison
How @pleach/core compares to the AI SDK, LangChain, LlamaIndex, Mastra, and to agent harnesses.
Adoption paths
Three phased on-ramps — from one-file prototype to multi-tenant production.
Upgrading to @pleach/core
When `simpleChatbot` and `createPleachRoute` aren't enough — direct `SessionRuntime` use, custom plugins, custom tools.
Ownership boundaries
Per SKU, what the package owns, what you configure, and what you must supply. The single-page map of where the substrate ends and your code begins.