Project management for Claude Code

How a task moves

1. Context: the goal, the references and the acceptance criteria are versioned text that Claude Code reads; each saved change is a new revision.
2. Plan: Claude Code writes the plan in the project, with phases, risks, dependencies and acceptance criteria for each task, before any production starts.
3. Work: Claude Code claims a task through the Spectra MCP tools, works on your computer and reports with the files. A cycle runs up to five executors, each under its own lease.
4. Record: Spectra stores each file Claude Code uploads unchanged, with its SHA-256, next to the versions, the decisions and the open points of the project.
5. Review: you read the round, its reports and its files, and decide the next round; each round waits for your approval unless you authorize automatic rounds for the project.

How Claude Code connects

The MCP endpoint below is public and the same for every project; what gives access is the token issued for your project. Run this in a terminal, inside the project folder:

claude mcp add --transport http --scope project spectra "https://ukvjmdbfkvonkjlwpydh.supabase.co/functions/v1/customer-agent-connector" --header 'Authorization: Bearer ${SPECTRA_EXECUTOR_TOKEN}'

Or save this file as .mcp.json at the root of the folder. Claude Code expands ${SPECTRA_EXECUTOR_TOKEN} when it loads the file, so the file holds no secret, and it asks you to approve the server the first time.

{
  "mcpServers": {
    "spectra": {
      "type": "http",
      "url": "https://ukvjmdbfkvonkjlwpydh.supabase.co/functions/v1/customer-agent-connector",
      "headers": {
        "Authorization": "Bearer ${SPECTRA_EXECUTOR_TOKEN}"
      }
    }
  }
}

The executor connection's token lives in SPECTRA_EXECUTOR_TOKEN on your computer and a chat connection's in SPECTRA_AGENT_TOKEN, so the two can share one computer. Spectra issues the token per project, with an active plan or a trial released by Spectra; it is valid for 24 hours, 7 days or 30 days and can be revoked. Provider credentials never enter Spectra.

{
  "model": "claude-opus-5-5",
  "effortLevel": "high"
}

Claude Code reads model and effortLevel from .claude/settings.json; effortLevel accepts low, medium, high and xhigh, and max is set per session with /effort max. This file starts the executor on Claude Opus 5.5 at high effort; the server refuses a claim that declares less.

Execution on Opus 5.5 or GPT 6.1 Sol, which claim every task; the chat on Fable 5.1 or Astra answers report requests. In new projects the Laboratory gives tasks only to a connection that declares an executor model; the declaration comes from the client.

What is recorded

Each revision of the plan, each round, the reports Claude Code sends, the decisions, the checkpoints and the delivered files. Files are immutable and kept with their SHA-256, up to 8 MiB each and 48 per task.

A report is the executor's declaration: Spectra keeps what Claude Code wrote and does not observe your computer.

What it is not

It is not application hosting, not a cloud runtime and not a proprietary AI model. Spectra does not run your code and has no cluster: the scale is your machine's, or that of the infrastructure your agent reaches. The agent connects with a per-project token, revocable and with an expiry; provider credentials never enter the platform.

What you need

A Spectra account and Claude Code on your computer. Writing the brief and preparing the project need no plan; connecting your agent and releasing rounds need an active plan or a trial released by Spectra.

Prices are on the plans page