Documentation

Everything you need to know
about Konklave

Complete reference for agents, modes, commands, and configuration.

Overview

Konklave is a council of AI models that work together as your full-stack development team — from planning to production — all through a single terminal command.

Konklave (Latin for "assembly" or "council") brings together a crew of specialized AI agents — Architect, Coder, Reviewer, Tester, Researcher, and Auditor — each powered by a best-in-class model. You describe what you need in plain language, and Konklave orchestrates the agents through a structured pipeline to deliver production-ready code, tests, and documentation.

Releases June 15, 2026. Konklave was built to solve a simple problem: AI is incredibly capable, but coordinating multiple models, managing context, and structuring output is still hard. Konklave wraps that complexity into a single command — konklave run "build me a REST API" — and handles the rest.

How It Works

Konklave routes your request through a multi-agent pipeline. Each agent is a specialized AI model that handles one phase of the development lifecycle.

konklave@docs:~
┌────────────────┐ ┌──────────┐ ┌───────────┐ ┌─────────┐ ┌────────┐ │ Architect │ ──→ │ Coder │ ──→ │ Reviewer │ ──→ │ Tester │ ──→ │ Result │ │ (plan + spawn) │ │ (write) │ │ (review) │ │ (test) │ │ ✅ │ └───────┬────────┘ └──────────┘ └─────┬─────┘ └─────────┘ └────────┘ │ │ │ spawns 3× │ consensus with ▼ ▼ Architect ┌─────────────┐ │ Researcher │ │ × 3 swarm │ │ (parallel) │ └──────┬──────┘ │ each returns ▼ ┌─────────────┐ │ Results │ │ to Architect│ └─────────────┘

The Architect receives your task first. It analyzes the request, breaks it down into subtasks, and produces a structured plan. The Coder then implements each subtask, writing real files to the workspace. The Reviewer and Auditor inspect the code in parallel — the Reviewer checks for bugs and style, the Auditor checks for security and compliance. The Tester writes and runs tests. If the pipeline needs external information — docs, API specs, package details — the Researcher fetches it.

Each agent passes its output to the next in a configurable sequence. The whole process streams to your terminal in real time, so you see every decision, every file written, and every test result as it happens.

Six Agent Roles

Every role in Konklave is backed by a different model, chosen for its strengths. This is the default mapping — you can override models in config.toml.

Role Model What It Does Cost
Architect Claude Sonnet 4.6 Designs the plan, splits work into subtasks, specifies interfaces, and guides the team through the pipeline. Medium
Coder DeepSeek Chat V3 Writes production-ready code for every module. Creates files, implements functions, and follows the architect's spec. Low
Reviewer Claude Sonnet 4.6 Catches bugs, enforces code style, checks for edge cases, and suggests concrete improvements. Blocks bad code from passing. Medium
Tester Claude Haiku 4.5 Writes unit and integration tests, runs them, and reports failures. Ensures every function is covered. Low
Researcher Claude Haiku 4.5 Looks up facts, documentation, API references, and package details that other agents need during the task. Low
Auditor Claude Sonnet 4.6 Strict final quality gate — reviews code for security vulnerabilities and compliance before the output is delivered. Medium

Work Modes

Konklave offers 7 built-in work modes. Each one trades off speed, depth, and autonomy differently. Choose the mode that fits your current task.

Mode Speed Description Pipeline
Very Fast fast No research → code → done. Fastest path to output. Great for quick experiments. quick · --work very_fast
Fast fast One researcher, quick review. Lightweight, skips test for speed. fast · --work fast
Balanced medium Research → plan → code → review → test. Full pipeline, default mode. build_feature · --work mid
Deep medium Multi-researcher, multiple review and test rounds. Deeper quality. build_feature (×1.5) · --work long
Exhaustive slow Exhaustive edge cases, trade-off exploration. Maximum depth with multi-phase planning. thorough · --work very_long
Autonomous slow Loop plan→code→review→test→audit until Auditor approves (budget-capped). autonomous_loop · --work auto
Auto-endless slow Same as autonomous but runs until you stop it (no budget cap). autonomous_loop · --work auto_endless

Pipelines

A pipeline is a sequence of agent roles that execute one after another. Konklave ships with these built-in pipelines:

build_feature

Research + plan + code + review + test + audit. Full feature development cycle.

quick

One-shot coder, no review overhead. Fastest possible path — ideal for trivial changes or simple scripts.

fast

Lightweight: one researcher, quick review, skips test. Good for moderate confidence.

thorough

Multi-researcher, divergent solution approaches, consensus review, testing, and strict audit.

autonomous_loop

Loop until done (budget-capped). Full agent team in iterative cycles, checking quality and re-working until gates pass.

autonomous_slice

Thin slice — dissent engine + northstar + one audit checkpoint. Quick autonomous run for small tasks.

You can define custom pipelines as .yaml files — place them alongside the built-in pipelines to extend Konklave with your own agent sequences.

CLI Commands

Konklave ships with a small but powerful CLI. Here are all the commands at your disposal:

$ konklave init

Run the interactive setup wizard — creates ~/.konklave/config.toml, stores your OpenRouter API key in the OS keyring, and configures default models for each role.

$ konklave run <task>

Run a task with your Konklave agent team. <task> is a natural-language description. Use --work / -m to select a work mode (default: mid — balanced).

$ konklave

Run without arguments to launch the interactive Textual TUI (Terminal User Interface).

$ konklave config

Show current configuration as a Rich table in the terminal. Use flags like --set-locale, --reset-api-key, --set-max-history-tokens to modify specific values.

$ konklave ping

Check connectivity to your configured AI providers. Pings each model and reports latency and status.

$ konklave resume

Resume the last interrupted session. Useful after a crash or when you want to continue from where you left off.

$ konklave tutorial

Launch an interactive tutorial that walks you through your first Konklave session step by step.

TUI Features

Konklave's Textual Terminal User Interface gives you a rich, interactive experience inside the terminal. Here's what you can do:

Feature Description Keyboard
Agent Panel See all six agents with their status — idle, thinking, working, waiting_human, blocked, done. Tab/Shift+Tab navigates between agent cards. Tab / Shift+Tab
Advanced View Full TUI with pipeline graph, agent cards, logs, and cost tracker. Ctrl+A
Simple View Minimal view showing only the essential agent status and output. Ctrl+S
Help Screen Full list of keyboard shortcuts and UI tips. F1
Copy Last Output (50 lines) Copy the last 50 lines of output to clipboard. F2
Copy Last Output (200 lines) Copy the last 200 lines of output to clipboard. F3
Slash-Command Mode Enter commands like :help, :cost, :commit directly. :
Multi-line Editor Open a multi-line input editor for composing prompts. Ctrl+E
Quit Exit the TUI cleanly. Ctrl+Q

Swarm Mode

Swarm mode controls how aggressively Konklave spawns sub-agents to parallelize work. It's a key lever for trading cost against speed.

Aggressive

Maximum parallelism. Spawns sub-agents freely to complete independent work simultaneously. Fastest but highest cost. Great for large codebases with many independent modules.

Encouraged

Balanced approach. Sub-agents are spawned when there's clear independent work, but the orchestrator doesn't over-split. The default mode — good for most tasks.

Off

Fully sequential. No sub-agents. Every phase runs in strict order. Lowest cost but slowest. Best for simple tasks or when you want maximum predictability.

Set swarm mode in config.toml with the swarm_mode key: "aggressive", "encouraged", or "off".

Configuration

Konklave is configured through a config.toml file at ~/.konklave/config.toml. Sensitive values (API keys) are stored in your OS keyring, not in the config file.

# ~/.konklave/config.toml
work_mode = "mid"
swarm_mode = "encouraged"
[default_models]
architect = "anthropic/claude-sonnet-4.6"
coder = "deepseek/deepseek-chat-v3"
reviewer = "anthropic/claude-sonnet-4.6"
tester = "anthropic/claude-haiku-4.5"
researcher = "anthropic/claude-haiku-4.5"
auditor = "anthropic/claude-sonnet-4.6"
# API key stored in OS keyring, not in this file

API Key Management. Your OpenRouter API key is stored securely in your operating system's keyring (Keychain on macOS, Credential Manager on Windows, Secret Service on Linux). During konklave init you will be prompted to store your OpenRouter API key in the OS keyring. The config file never contains secrets.

Model Overrides. You can override any role's model in config.toml. Set a different model name per role — swap any agent to a different model via the default_models section.

Work Mode. The work_mode key sets the default pipeline. Override per-run with konklave run --work <name>.

Installation

$ pip install konklave
Python 3.11+ OpenRouter API key MIT License

After installing, run konklave init in your project directory to create a default config file and store your OpenRouter API key.

On Windows, run konklave in Command Prompt, PowerShell, or Windows Terminal. On macOS/Linux, just run konklave in your terminal.

Platform Support

Konklave runs on all major desktop platforms. Here's what you need:

🪟 Windows

Windows 10 or 11. Python 3.11+ via the official installer or winget. Full TUI support via Windows Terminal.

🍎 macOS

macOS 13+ (Ventura) or later. Python 3.11+ via Homebrew or the official installer. Works with both Terminal.app and iTerm2.

🐧 Linux

Any modern Linux distribution with Python 3.11+. Tested on Ubuntu 22.04+, Fedora 38+, and Arch Linux. Install via pip. TUI requires a terminal emulator with true-color support.

License

Konklave is released under the MIT License — free to use, modify, and distribute. No strings attached.

The full license text is included with every distribution in the LICENSE file, and also available in the GitHub repository.

MIT License