SpecStrand Docs

Documentation

SpecStrand turns ideas into reviewable, agent-executable plans and tasks, and drives them with an autonomous work loop. These pages are the reference for the spec CLI, configuration, and connecting an agent over MCP.

New here? Start with the getting-started guide: install, authenticate, and run your first task in a few minutes. Just after the binary? Go straight to installing the CLI.

Concepts: graphs first

Work is organized from the top down. A graph holds related plans and the order they drain in. A plan is a named set of tasks, and a task is one unit an agent claims, works, and completes. A project groups plans for scoping.

  • Start with the graph. At the start of multi-plan work, an agent lists the graphs (graph.list, spec graph list) and reads the one that fits; if none fits, it creates one before the first plan.
  • Create plans into the graph (plan.create with graph, or spec plan import --graph G). A plan can stay outside every graph, but each create or update then returns a plan_not_in_graph warning, and the plan page shows Not in a graph with an Add to graph action.
  • Dependencies between plans set the drain order; graph.get shows which plans are ready now and what blocks the rest.

The Getting started graph

A new account starts with one real graph, Getting started, written for your agent. You connect an agent over MCP, tell it to work the graph, and watch it drive the loop in Studio while a guide line under the app bar points at what to look at.

  • The graph lives in its own project, Getting started (getting-started), so the agent drives it with project.work and that project and never touches your real work. If you already have a project with that slug, the graph gets a new one (getting-started-2) and leaves yours alone; the prompt in the app names the right one.
  • It confirms who it is, asks you for one small idea, turns the idea into a plan inside the graph, and works that plan with task.report and task.complete.
  • A gate stops the agent before the last part. Only you can answer it, in Studio; the agent cannot bypass it.
  • The last part proposes a SpecStrand section for your project's AGENTS.md (or the rules file your agent reads). The agent asks you first in its own window and writes nothing unless you say yes.
  • Delete the graph from its tab in Studio when you are done. Deleting archives its template plans, so it can come back; your idea and the plan your agent made stay.
  • An account created before the graph existed can add it from Settings → MCP.

Reference

  • Install the CLI: the one-line installers, direct downloads with checksums, and the supported platforms.
  • CLI command reference: every spec command, generated from the binary's own help, plus the exit-code contract.
  • Configuration and environment: the CLI environment variables and workspace configuration.
  • MCP setup: wire the SpecStrand MCP server into an MCP-capable agent and drive the loop with structured tools.