"Where no AI is King, but all serve the Code."

A CLI tool that lets multiple AI models (Claude, Gemini, GPT) automatically discuss software problems and reach consensus — so you don't have to be the messenger between them.

1. You ask a question
2. TheRoundtAIble sends it to your configured AI "knights"
3. Each knight responds with their proposal and scores agreement (0-10)
4. Rounds continue until consensus is reached or max rounds hit
5. You choose: let the Lead Knight execute, do it yourself, or decide later

You: roundtable discuss "How should we refactor the auth module?"

  ROUND 1 — KNIGHTS! DRAW YOUR KEYBOARDS!

    Claude (7/10) → proposes JWT with NextAuth wrapper
    Gemini (8/10) → agrees, suggests refresh token strategy
    GPT    (6/10) → "Can we just ship it?"

  ROUND 2 — FOR KING AND KONG!

    Claude (9/10) → accepts refresh strategy
    Gemini (10/10) → fully agrees
    GPT    (9/10)  → "Fine, let's do it properly"

  Against all odds... they actually agree.

  What is your decree, Your Majesty?
  1. Let the knights forge it — they write the code
  2. I'll wield the sword myself — just show me the plan
  3. Adjourn the court — decide later

No API keys required by default. Works with your existing AI subscriptions (Claude Pro/Max, Gemini Advanced, ChatGPT Plus). API keys are an optional fallback.

npm install -g theroundtaible

Or use without installing:

npx theroundtaible init

npm update -g theroundtaible

Check your version:

roundtable --version

• Node.js 20+
• At least one AI tool:
  • Claude Code — claude command
  • Gemini CLI — gemini command
  • Codex CLI — codex command
  • OpenAI API key (OPENAI_API_KEY env var) as fallback
  • Ollama or LM Studio for local models

# Initialize in your project (run this once per project)
cd your-project
roundtable init

# Start a discussion
roundtable discuss "How should we structure the database schema?"

# Review based on current git changes
roundtable summon

# Apply the consensus decision
roundtable apply

# Emergency bug diagnosis
roundtable code-red "login page crashes on submit"

Note: Each project needs its own roundtable init. This creates a .roundtable/ folder with project-specific config, chronicle (decision history), and session data. Your API keys are stored globally (~/.theroundtaible/keys.json), so you only enter them once.

After roundtable init, your .roundtable/config.json controls everything:

{
  "knights": [
    {
      "name": "Claude",
      "adapter": "claude-cli",
      "capabilities": ["architecture", "debugging", "testing"],
      "priority": 1,
      "fallback": "claude-api"
    },
    {
      "name": "Gemini",
      "adapter": "gemini-cli",
      "capabilities": ["docs", "ui-ux", "planning"],
      "priority": 2
    }
  ],
  "rules": {
    "max_rounds": 5,
    "consensus_threshold": 9,
    "timeout_per_turn_seconds": 120
  }
}

Consensus threshold: All knights must score >= this value AND have no pending issues.

Code-Red is temporarily disabled while we focus on making discuss and apply rock-solid. The diagnostic code is written and will return in a future update. For now, use roundtable discuss "describe the bug" — the knights are surprisingly good at debugging through discussion.

When a bug won't die, call in the doctors:

roundtable code-red "the API returns 500 on user creation"

The knights switch to diagnostic mode:

• Triage round: Initial assessment of symptoms
• Blind round: Each doctor diagnoses independently (prevents groupthink)
• Convergence rounds: Doctors compare findings and converge on root cause
• File requests: Doctors can request specific source files as evidence

When they agree on a diagnosis, you choose: Fix now, Report only, or Log for later.

TheRoundtAIble tracks what has been built. After each roundtable apply, the manifest is automatically updated so knights don't re-propose features that already exist.

# See what's been built
roundtable manifest list

  [+] auth-refactor — JWT with NextAuth wrapper
      Status: implemented | Knight: Claude | 2026-02-17
      Files: src/lib/auth.ts, src/middleware.ts

  [~] user-dashboard — Dashboard component with charts
      Status: partial | Knight: Gemini | 2026-02-17
      Files: src/components/Dashboard.tsx
      Skipped: src/components/Charts.tsx

# Check for stale entries (files that no longer exist)
roundtable manifest check

Knights see the manifest in their system prompt, preventing duplicate proposals.

When knights reach consensus (score >= 9), they declare which files they intend to modify. The apply command enforces this scope:

• Files outside the agreed scope are blocked (shown in red)
• Use --override-scope to bypass (requires typing "YES" + a reason, logged for audit)
• Old sessions without scope data work normally (no enforcement)

your-project/
  .roundtable/
    config.json          # Which knights, rules, capabilities
    chronicle.md         # Decision log (persistent memory)
    manifest.json        # Implementation manifest (what's been built)
    error-log.md         # Code-red diagnostic log
    sessions/
      2026-02-16-auth-refactor/
        topic.md         # The question
        discussion.md    # Full discussion (all rounds)
        decisions.md     # The consensus decision
        status.json      # Current phase and progress

See architecture-docs.md for the full technical architecture, project structure, and contribution guide.

• Multi-AI discussions with consensus
• Claude, Gemini, and OpenAI adapters
• Knight personalities (they roast each other)
• Apply decisions with Parley/No Parley modes
• Code-Red emergency diagnostic mode (deferred to v1.1 — focusing on discuss + apply first)
• Git-diff based discussions (summon)
• Chronicle (persistent decision memory)
• Implementation Manifest (tracks what's been built)
• Scoped Apply (knights declare files, apply enforces scope)
• Local LLM support (Ollama, LM Studio)
• VS Code extension
• Web dashboard for session visualization
• More adapters (DeepSeek, Llama, Mistral)
• CI/CD integration (GitHub Actions)

TheRoundtAIble supports local models through Ollama and LM Studio. Run roundtable init and any running local server will be auto-detected.

We recommend Ollama. Here's why:

Ollama gives TheRoundtAIble full programmatic control — context window size is detected automatically and allocated dynamically per prompt. LM Studio requires you to manually configure Context Length and Response Limit in the GUI.

Local models run on your GPU. Bigger models = better discussion quality, but more VRAM:

Our honest take: Models under 30B struggle with multi-round discussions. They tend to repeat themselves, ignore other knights' arguments, and miss the collaborative spirit (no roasting!). Cloud knights (Claude, Gemini, GPT) handle 100K+ tokens and produce richer debates. Local models shine when you want privacy or offline access — but for best results, mix them with at least one cloud knight.

If you choose LM Studio, you must manually adjust these settings (Developer tab > Model Settings):

• Context Length: increase to at least 16384 (default 4096 is too small)
• Response Limit: uncheck the limit, or set to 4096+

Higher context = more VRAM and slower responses. Find the sweet spot for your GPU.

• Ollama: Uses the native /api/chat endpoint with dynamic num_ctx — only allocates as much context as the prompt needs, saving GPU memory
• LM Studio: Uses the OpenAI-compatible /v1/chat/completions endpoint
• Auto-detection: roundtable init probes localhost:11434 (Ollama) and localhost:1234 (LM Studio), discovers loaded models, and filters out non-chat models (embeddings, TTS, etc.)
• Context budgeting: The orchestrator detects each local model's context window and adjusts the source code payload so it fits — cloud knights get the full context, local knights get a trimmed version

roundtable apply works but is in beta. The block-level system (RTDIFF/1) lets knights target specific functions and classes instead of rewriting entire files, which solved the reliability issues with large files. Validation catches bad output before anything is written — your code is safe.

What works:

• roundtable discuss — multi-AI discussions work reliably
• roundtable apply — writes code via block-level operations (tested on 600+ line files)
• roundtable apply --dry-run — test the full pipeline without writing anything
• Validation pipeline — blocks bad output (157 tests, no corrupted files)
• Scope enforcement + backup system

Known risks:

• Output quality depends on the LLM — validation blocks bad code but nothing gets written
• Single attempt, no retry — if the knight fails, you re-run or apply manually
• More real-world testing needed

If apply fails, read the decision in .roundtable/sessions/*/decisions.md and apply manually.

MIT — Tarik Polat (polatinos)