Ploston

LLMs plan. Ploston executes.

Ploston is a deterministic execution layer for AI agents. You define multi-step workflows in YAML, Ploston exposes them as MCP tools, and your agent calls them with a single invocation.

---

The problem

When an agent orchestrates its own tool calls, every intermediate result passes through the LLM. A 3-step task burns ~10,000 tokens. Run it twice and you may get different results. When something fails, you get a reasoning blob with no trace.

The solution

Ploston moves orchestration out of the LLM. Same 3-step task: ~500 tokens. Same inputs → same outputs, every time. Full step-by-step trace.

Why this matters →

---

Quick start

pip install ploston-cli       # install the CLI
ploston bootstrap             # deploy the Control Plane (Docker)
ploston init --import         # import your MCP tools, start local runner
# restart Claude Desktop

Full installation guide →

---

How it works

┌──────────────┐  MCP (stdio/HTTP)  ┌───────────────────────────┐
│ Claude / any │ ◄────────────────► │  Ploston Control Plane    │
│  MCP client  │                    │  ┌──────────┐ ┌─────────┐ │
└──────────────┘                    │  │ Workflow │ │  Tool   │ │
                                    │  │ Engine   │ │Registry │ │
                                    │  └────┬─────┘ └────┬────┘ │
                                    └───────┼────────────┼──────┘
                                            │  WebSocket │
                                    ┌───────▼────────────▼──────┐
                                    │       Local Runner        │
                                    │  github, postgres, slack, │
                                    │  your custom MCP servers  │
                                    └───────────────────────────┘

Control Plane — workflow engine, MCP frontend, tool registry, REST API. Runs in Docker.

Local Runner — connects your local MCP servers to the CP over WebSocket. Your tools never leave your machine.

Workflows — YAML files. Once registered, appear as w_<name> MCP tools.

---

Write a workflow

name: page-report
version: "1.0.0"
description: "Fetch a URL and save a title report"

inputs:
  - name: url
    type: string

steps:
  - id: fetch
    tool: http_request
    params:
      url: "{{ inputs.url }}"
      method: GET

  - id: extract
    depends_on: [fetch]
    code: |
      import re
      body  = context.steps["fetch"].output.get("body", "")
      match = re.search(r"<title[^>]*>(.*?)</title>", body, re.IGNORECASE)
      result = {"title": match.group(1).strip() if match else "No title"}

  - id: save
    depends_on: [extract]
    tool: fs_write
    params:
      path: "report.txt"
      content: "Title: {{ steps.extract.output.title }}\n"

outputs:
  - name: title
    from: steps.extract.output.title

ploston validate page-report.yaml   # check it
ploston run page-report.yaml -i url=https://example.com   # test it

Claude can now call w_page-report as a single MCP tool.

First workflow tutorial →

---

What's included

	OSS (available now)	Enterprise (planned)
Workflow engine	✅	✅
MCP integration	✅	✅
Python sandbox	✅	✅
CLI + REST API	✅	✅
Docker + K8s deploy	✅	✅
Prometheus + Grafana	✅	✅
RBAC / policy engine	—	Phase 4
Parallel execution	—	Phase 4
Pattern mining	—	Phase 5
Workflow synthesis	—	Phase 5

Full roadmap →

---

Documentation

Getting started → Installation — deploy in 5 minutes → Quickstart — write and call your first workflow → First Workflow Tutorial — step-by-step with tool calls

Concepts → How Ploston Works — planning vs execution separation → Execution Model — steps, data flow, error handling → Security Model — 7-layer sandbox → Workflows as Tools — MCP tool publishing

Guides → Workflow Authoring → Code Steps → Tool Integration → Docker Deployment → Troubleshooting

Reference → CLI Reference → Workflow Schema → Configuration → Error Codes

---

System requirements

• Python 3.12+
• Docker & Docker Compose v2.20+
• macOS, Linux, or Windows (WSL2)
• 2 GB RAM recommended

License

Apache 2.0