Open source · MIT · macOS, Windows, Linux

Your AI,
organised into a team.

Eleven roles, one loop, shared files. Drop one folder into any project and your coding agent works like a full software team: product, design, architecture, engineering, QA, DevOps, security and docs, led by a CTO.

Install in a minute Star on GitHub Open the live UI

11 roles1 loop0 servers

Fig. 1Eleven strings, one loop. M-theory needs eleven dimensions to unify five string theories. eleven unifies eleven roles into one team.

01Install in a minute

One folder.
Invisible to git.

eleven installs to ./eleven in your project and hides itself with .git/info/exclude, so your repo and your teammates never see it. No dependencies, no server.

  1. Clone the kit once

    One copy on your machine serves every project.

  2. Install into a project

    Copies eleven/ in and hides it from git. Run it again any time; --update never touches your work.

  3. Tell your agent to lead

    Claude Code, Codex, opencode, Cursor: anything that reads AGENTS.md.


      

      

        
Then tell Claude Code, Codex, opencode or Cursor:

        Read eleven/AGENTS.md and act as the CTO. New item 001: <your idea>. Start at step 0.
        
      

    

  







  

    

      
02How it works

      
A loop,
not a long chat.

    

    
Long chats rot: context fills with noise and the model forgets its own decisions. eleven gives each task a fresh agent and keeps memory in a small, rewritten set of files called the ledger.

  

  

  
    
    
  1. 01
    Intake
    product
    Product writes a brief with testable criteria. You approve.
    Gate
    2. 
    
  2. 02
    Plan
    cto
    The CTO writes a short strategy and seed tasks.
    3. 
    
  3. 03
    Explore
    architect
    The architect maps the hard part, the options and the pitfalls. No code yet.
    4. 
    
  4. 04
    Manage
    cto
    The CTO picks exactly one next task for one role.
    5. 
    
  5. 05
    Execute
    any role
    A fresh agent does that task and files a report with evidence.
    6. 
    
  6. 06
    Verify
    qa
    QA checks every criterion. A red check beats any "done".
    7. 
    
  7. 07
    Ship
    cto
    Merge and release small and often. You accept.
    Gate
    8. 
  

  

    
Guards

    
    
      
  • 10 round budget
  • No-progress stop
  • plan.md ≤ 4k chars
  • notes.md ≤ 8k chars, rewritten
  • Open questions pause for you
  • One task per fresh context
    • 
    

  







  

    

      
03The team

      
Eleven roles.
You're the CEO.

    

    
Each role is one short Markdown file: what it does, what it never does, and when it's done. Not every task needs everyone. A bug fix is just CTO → engineer → QA.

  

  

    

      
      
    

    
    
      
  • CTOlead
    Runs the loop. Plans, picks the next task and role, merges, ships, and enforces the guards.
    • 
      
  • Product manager
    Brief, scope, non-goals, testable acceptance criteria.
    • 
      
  • Designer
    Flows, screens, every state, real copy, accessibility.
    • 
      
  • Architect
    The core difficulty, the approaches, interfaces and decision records.
    • 
      
  • Backend
    APIs, domain logic, data, safe migrations. Test-first.
    • 
      
  • Frontend
    Web UI built to design in every state, accessible and fast.
    • 
      
  • Mobile
    iOS and Android: offline, app lifecycle, permissions.
    • 
      
  • QA
    Verifies every criterion with evidence. Can veto "done".
    • 
      
  • DevOps
    CI, deploys, rollback, observability.
    • 
      
  • Security
    Threat models and reviews. High severity blocks shipping.
    • 
      
  • Tech writer
    README, docs and changelog that match what shipped.
    • 
      
  • YouCEO
    Approve the brief, answer questions, accept the result.
    • 
    

  







  

    

      
04The live UI

      
See your team
work.

    

    
Open your project folder in the browser. The page reads the ledger while your agents write it. Files stay on your computer.

  

  

    

      
The work graph: the project, its items, each item's tasks coloured by status, and the reports, design files and decisions they producedThe work graph: the project, its items, each item's tasks coloured by status, and the reports, design files and decisions they produced

      
Home: one card per item; the item waiting on you comes first, in amber, with its open questions foldedHome: one card per item; the item waiting on you comes first, in amber, with its open questions folded

      
The reader: the file tree, a report open in the middle, and the work graph beside itThe reader: the file tree, a report open in the middle, and the work graph beside it

    

  

  

    
1
Project → items → tasks → outputs
The work graph shows every task, coloured by status, with the reports, designs and decisions it produced. Click any node to open it.

    
2
What's happening, what waits on you
One card per item. Items that wait on you come first, in amber, with their open questions folded away.

    
3
Read it like Obsidian
A file tree on the left, the doc in the middle, the graph beside it. Paths in a doc are links. Light and dark.

  

  
Open the live UI







  

    

      
05Built on research

      
Organisation beats
raw model size.

    

    
eleven's loop comes from GVS5H (Gao et al., 2026): one model, split into a manager and fresh single-task workers that coordinate through files on disk.

  

  

    

      
+25.6

      
points on hard LiveCodeBench problems for Qwen3.8-27B.

      
    

    

      
93.0vs90.4

      
An orchestrated cheap model beat a frontier model…

      
    

    

      
~9%

      
…at about 9% of the cost.

      
    

  

  

    
    
      
  1. Decomposition: one task per call keeps every agent focused.
    2. 
      
  2. Persistent context: the ledger outlives any single conversation.
    3. 
      
  3. Reasoning on disk before code: notes first, then implementation.
    4. 
      
  4. Verifier override: a failing test beats a confident "done".
    5. 
    

    

      
To be fair about itThe scaffold uses about 3× the tokens of a single call, strong models gain least, and some configurations got worse. eleven adds the roles a real product needs (design, security, DevOps, docs) and keeps you at every gate.

      
Source: Gao et al., 2026, arXiv:2608.26480

    

  







  

    

      
06Why eleven

      
Small enough to read.
Strict enough to trust.

    

    
The whole kit is about 25 Markdown files. Read it in fifteen minutes, then change anything.

  

  

    
AGENTS.md
Works with your agent
Claude Code, Codex, opencode, Cursor: anything that reads AGENTS.md. No lock-in, no framework.

    
/app
Live UI in the browser
Open your project at eleven.rtima.org/app or locally: a file tree, a reader for every doc, and a work graph. It updates every 2 seconds and nothing is uploaded.

    
ui/ → Home
Waiting on you, first
When the next step is yours, that item's card comes first, in amber: one line on what to do, with the open questions folded underneath.

    
reports/
Evidence, not vibes
Every report pastes the real build and test output. QA re-runs it.

    
decisions/
Decisions on record
Expensive choices get a one-page record, so nobody re-argues them next month.

    
install.sh
install.ps1
Cross-platform
install.sh for macOS and Linux, install.ps1 for Windows. Safe re-installs; --update never touches your work.

  







  

    
07FAQ

    
Questions,
answered.

  

  

    
Q1
Does it cost more tokens?
Yes, often about 3× a single chat for a full loop. You buy fewer reworks, fewer surprises and a written trail of why. For tiny fixes, skip roles: CTO → engineer → QA.

    
Q2
Does my code or ledger go anywhere?
No. eleven is just files in your project. The live UI reads your folder in your browser through the File System Access API; nothing is uploaded.

    
Q3
Which browsers support the live UI?
Chrome and Edge give live updates. Firefox and Safari open a snapshot; pick the folder again to refresh it. Either way the files stay in your browser and nothing is uploaded. HTML files from the ledger open in a sandboxed frame.

    
Q4
Will my teammates see it?
No. It's excluded through .git/info/exclude, which is local to your clone. To share the ledger with your team, remove that line and commit eleven/work.

    
Q5
Can I change the roles?
Please do. When an agent repeats a mistake, add a line to its role file. That's how your team learns.

  







  
  

    
Stop chatting.
Start leading.

    
Give your agent a team, a loop and a definition of done.

    

      Install eleven
      Open the live UI