councli

councli is a local council room for native coding-agent CLIs. It broadcasts a prompt to installed assistants, lets them see each other's output through a shared blackboard, and returns one synthesized answer.

It does not replace Codex, Claude Code, AGY, CodeWhale, or Kimi Code. Those tools remain normal binaries on your machine. councli is the coordination layer around them: room, recorder, router, and shared memory.

Positioning

Multi-model platforms and agent frameworks already exist. Some route prompts to hosted models, some synthesize model panels, and some help developers build new agent workflows.

councli is narrower: it coordinates the coding assistants you already run in your terminal. It preserves their native harnesses, auth, tools, permissions, slash commands, MCP configuration, and session behavior instead of rebuilding them as API agents.

In short: councli is the missing local room for developers who already use multiple coding assistants and want them to think together inside the same repo.

Install

pipx install councilroom-ai
councli --help

The PyPI package is councilroom-ai; the installed command is councli.

From GitHub:

pipx install "git+https://github.com/mosesmrima/councli.git"

Supported assistants

The MVP supports these coding-agent CLIs:

Assistant	Command
Codex	codex
Claude Code	claude
AGY	agy
CodeWhale	codewhale
Kimi Code	kimi

Each assistant must already be installed, logged in, model-ready, and available on PATH in the same shell where councli runs.

councli does not manage provider accounts, subscriptions, API keys, OAuth, device-code login, or model configuration. Launch each assistant directly first, finish its native setup, then let councli discover and coordinate it.

What it does

• Runs shared multi-agent conversation turns from one prompt.
• Records all participant outputs on an inspectable blackboard.
• Synthesizes one council answer from the shared outputs.
• Supports explicit /deliberate and /vote commands when you want stronger coordination.
• Lets you attach to a native assistant TUI with /assistant <name> when tmux is available.
• Tracks artifacts under .councli/ instead of hiding coordination state.
• Degrades gracefully when an assistant is missing, disabled, unauthenticated, or missing a model.

What it is not

• Not a model provider.
• Not an auth manager.
• Not a replacement for the underlying assistants.
• Not a generic sandbox.
• Not a fixed workflow engine that forces every prompt through vote/review phases.

Normal prompts are just shared conversation turns. Stronger governance is explicit.

Quick start

From the project you want the assistants to inspect:

councli setup
councli doctor
councli

Inside the interactive shell:

councli > what can you all do?
councli > /deliberate compare sqlite and postgres for this app
councli > /vote choose the transport: exec or tmux
councli > /assistant codex
councli > /quit

Useful setup commands:

/agents
/enable claude
/disable kimi
/doctor
/status
/show latest

How a turn works

flowchart LR
    U([User prompt]) --> R{Council room}
    R -->|fan out| A1[Codex]
    R -->|fan out| A2[Claude Code]
    R -->|fan out| A3[AGY / CodeWhale / Kimi]
    A1 --> BB[(Shared blackboard)]
    A2 --> BB
    A3 --> BB
    BB -->|synthesize| S([One council answer])
    S --> U

Loading

Under the hood:

user prompt -> fan out to available assistants -> write responses to the
blackboard -> synthesize one council answer -> return to the prompt

/deliberate <prompt> adds a peer-aware second round before synthesis. /vote <prompt> asks participants for explicit decision artifacts.

Platform support

councli is a pure Python CLI for Python 3.11+.

Platform	Install	Core turns	Native tmux attach
Linux	Supported	Supported	Supported with tmux
macOS	Supported	Supported	Supported with tmux
Windows	Supported	Supported with exec-mode agents	Use WSL for tmux
WSL	Supported	Supported	Supported with tmux

Safety model

councli launches the assistant binaries configured in .councli/config.yaml. Generated command templates and resolved binary paths are trusted before use. If commands, binary paths, hashes, backends, or enabled flags change, review the config and run:

councli trust

Run diagnostics:

councli doctor
councli doctor --json
councli security

Artifacts can contain prompts, responses, terminal output, and project context. Before sharing artifacts, use:

councli artifacts scrub --dry-run
councli artifacts export --output support.tar.gz

Roadmap

Current priorities:

• Better harness adapters for Codex, Claude Code, AGY, CodeWhale, and Kimi Code.
• Stronger shared communication protocol: turn packets, blackboard, sidecars, event log, and recovery.
• Better token efficiency: compact prompts, peer-context summaries, fewer automatic rounds, and artifact references instead of repeated context.
• Better interactive UX around slash commands, participant status, synthesis, and native attach.

See Roadmap.

Development

git clone https://github.com/mosesmrima/councli.git
cd councli
uv sync
uv run pytest -q
uv build

Documentation

• Install guide
• Architecture
• Protocol design
• Adapter contract
• Participant adapter notes
• Roadmap
• Terminal substrate
• Security model
• Packaging and release notes
• Research findings

Status

councli is early alpha software. The current public surface is shared conversation, deliberation, explicit voting, native attach, durable artifacts, and adapter readiness. Hidden experimental execution/review commands may exist behind COUNCLI_EXPERIMENTAL=1, but they are not the MVP path.

License: MIT.