Skip to main content

The Fibe SDK

The Fibe SDK is the interface to Fibe — a single Go binary that runs in three modes depending on who's calling:

  • CLI: you, at a terminal. fibe playgrounds list, fibe agents chat, fibe trick run.
  • Go library: programs you write in Go that embed Fibe directly. Automate deployments, build a custom dashboard, write a control plane.
  • MCP server: AI agents (Claude Code, Cursor, Antigravity, anything that speaks the Model Context Protocol). Agents call fibe_* tools and Fibe responds.

All three share the same authentication, the same resource model, the same retry / circuit-breaker / rate-limit logic. Pick your interface; the platform is the same.

When to use which mode

You are…Use…
At a terminalThe CLI: fibe ...
A Go program that needs to drive FibeThe Go library at github.com/fibegg/sdk/fibe
An LLM agent (Claude Code, Cursor, Codex, etc.)The MCP server: fibe mcp serve and 42 tools
Writing a one-off shell scriptThe CLI with -o json for parseable output
Building CI/CD automationThe CLI from a workflow, with an API key
Embedding Fibe in your own SaaSThe Go library

The 60-second tour

# Install (macOS / Linux)
brew install fibegg/sdk/fibe

# First-time setup
fibe login # browser device-code flow
fibe doctor # check connectivity + auth

# Daily use
fibe playgrounds list
fibe agents chat my-genie --text "Hello"
fibe tricks trigger --playspec-id 42 --from-file inputs.json

# Run an MCP server for your AI agent
fibe mcp serve
fibe mcp install --client claude-code

That's the whole product surface, in a paragraph.

What it talks to

The SDK is a client to the Fibe API. It doesn't run anything itself; it tells the platform what to do. On the other end is Fibe, which orchestrates Marquees (your Docker hosts), Props (your Git repos), Templates, Playgrounds, Tricks, and Genies.

So the SDK is what gets you those resources from a script, an agent, or a CI job — same as the web UI gets them from a browser.

The three modes in detail

CLI

fibe <resource> <action> [flags]

A standard command tree. Every top-level resource family has the usual list / get / create / update / delete plus action-specific subcommands. Output defaults to a readable table; -o json and -o yaml produce parseable output for scripts. Many commands accept -f file.json to load a payload from a file or - for stdin.

Read on: Install the CLI, Authentication, CLI reference.

Go library

import "github.com/fibegg/sdk/fibe"

client, _ := fibe.NewClient(fibe.WithAPIKey(os.Getenv("FIBE_API_KEY")))
pg, _ := client.Playgrounds.Create(ctx, &fibe.PlaygroundCreate{Name: "demo", PlayspecID: 5})

A thin Go wrapper over the Fibe HTTP API with sensible defaults: automatic retry on transient failures, circuit-breaker when the upstream is sick, idempotency-key generation for safe re-tries, rate-limit awareness, structured errors. Resource "managers" (client.Playgrounds, client.Agents, client.Tricks, …) mirror the REST shape.

Read on: Go library.

MCP server

fibe mcp serve                       # stdio, single-tenant
fibe mcp serve --transport sse # SSE, multi-tenant
fibe mcp install --client claude-code

The same Go binary doubles as an MCP server. AI agents call fibe_* tools (41 of them across resource CRUD, playground actions, agent control, greenfield setup, multi-step pipelines, monitoring, repo management, and a few escape hatches). The catalog is in Tools catalog and each tool has its own detail page under Reference → Tools.

Read on: MCP server, Tools catalog.

What's next