How Plico compares
Where Plico fits among TypeScript agent tools, and what it does not do yet.
Plico is a developer preview, and the field is crowded. This page says what Plico is for, what it does today, and when another tool is the better choice. It describes other projects only from their public documentation, and they move fast, so check theirs.
What Plico is for
Plico treats an agent as a folder of files you review like code: Markdown instructions and skills, TypeScript tools, config, evals. Around that folder it gives you the parts you would otherwise build yourself:
- One event log per run, which Studio, replay, evals and the HTTP API all read.
- Human approval that waits across a process restart.
- Studio, a local web UI to chat, inspect runs, decide approvals and review artifact versions.
- A versioned artifact workspace with diffs and export.
- Local SQLite with no account, gateway or hosted service.
- Embedding:
createPlicoruns inside your own backend and leaves authentication to you.
Where it sits
| Project | What it is | How Plico differs |
|---|---|---|
| Vercel AI SDK | A provider-agnostic toolkit for model calls, streaming and UI. | A library you call from code. Plico uses its provider packages and adds the project folder, run log and Studio on top. |
| Mastra | A TypeScript framework with agents, graph workflows, memory, MCP, evals and observability. | Mastra is broader today. Plico is narrower and file-first: the project is a folder, not a code-defined graph. |
| OpenAI Agents SDK | A lightweight SDK for multi-agent workflows with handoffs, guardrails and tracing. | Built around code-defined multi-agent flows. Plico runs one agent per project, with files as the contract. |
| LangGraph | A low-level orchestration framework for stateful agents. | Lower level and more flexible. Plico is higher level and opinionated about layout. |
| eve | Vercel's file-first agent framework, close to 1.0 when this was written. | The closest in spirit: an agent folder, tools, skills, durable sessions, approvals and evals. eve is further along. |
What Plico does not do yet
Plico's scope is deliberately small for 0.1. These are not in the packages today:
- MCP and external connections. Tools are TypeScript files in
tools/. - Agent memory. The
memory/folder is required in the layout, but the agent does not read or write it yet. - Multi-agent and workflows. One agent per project.
agents/,workflows/and theentryconfig key are reserved for later; see Project layout. - Context compaction. A long thread is sent to the model as it stands.
- Chat channels such as Slack, and a React component kit. Studio is the UI, and it is a local development tool, not something to deploy over production data.
- Hosted infrastructure. Plico runs where you run it. Persistence is SQLite on that machine.
The remaining plan lives in the GitHub milestones.
When to choose something else
- You need MCP servers, graph workflows or multi-agent handoffs now: look at Mastra, LangGraph or the OpenAI Agents SDK.
- You want chat channels such as Slack out of the box: look at eve.
- You only need model calls and streaming inside an app you already have: the AI SDK on its own is enough.
When Plico fits
- You want the agent reviewed in pull requests, with instructions and tools as files.
- You want to see exactly what a run did, replay it, and stop a risky tool call for a person.
- You want it local and self-hosted, with an HTTP API you can embed behind your own login.
Start with the Quickstart.