CCCC Coordination Console
Run coding agents as a durable, trackable team across runtimes, machines, and remote operations.
Per-dimension scores and reasoning
Evidence shows: README explicitly states Web UI is high-privilege and requires Admin Access Token; daemon IPC binds to localhost by default; IM tokens are read from environment variables; Group Bridge uses a trust model with explicit read/full authorization; capability allowlist governs MCP capabilities. This reflects least privilege. However, user confirmation is only mentioned for Group Bridge pairing and Web Access settings, not covering all high-risk operations, so user_confirmation scores 1. Data flow transparency: README describes ledger recording all messages and events, but does not detail data flow and storage locations, so scores 2. Sensitive data handling: IM tokens from env vars, runtime state in CCCC_HOME, but no mention of encryption, so scores 2. Dependency security: pyproject.toml lists dependency version ranges, CI has caching, but no vulnerability scanning evidence, so scores 2. External effects: Group Bridge allows remote operations but requires explicit authorization, and README warns against public exposure, so scores 2. Rollback: installer tests cover rollback, but no user-level rollback documentation, so scores 2. Source attribution: no contributor guide or author info, so scores 1.
Evidence shows: README and CI config are consistent, Python and Rust implementations coexist, but README clearly states Rust is experimental and Python is stable, reflecting self-consistency. Dependency availability: pyproject.toml lists dependencies, CI has caching, but no availability guarantee, so scores 2. Failure messages: README provides doctor command and FAQ, but no detailed error message examples, so scores 2.
Evidence shows: README defines target audience (multi-agent coordination) and multiple use cases, with a suitability table. Capability boundaries: README describes support for 17 runtimes and custom, and Group Bridge permission levels. Trigger precision: README describes message routing and automation rules, but no detailed trigger conditions, so scores 2. Environment fit: README provides multiple installation methods (pip, Docker, source) and supports multiple platforms, so scores 2.
Evidence shows: README is well-structured with quick start, architecture, CLI reference, etc., good information architecture. Install notes are detailed, including pip, Docker, source. Naming stability: version numbers consistent in Cargo.toml and pyproject.toml, but no naming convention doc, so scores 1. Examples and FAQ: README provides multiple examples and FAQ links, so scores 2. Known limitations: README mentions Rust implementation is experimental and some features unsupported, so scores 2. License: Apache-2.0 license file present, so scores 2. Versioning changelog: no CHANGELOG file, so scores 1. Maintenance responsibility: no maintainer info or contribution guide, so scores 1.
Evidence shows: README describes outputs (Web UI, CLI, MCP tools) and available features, output usability is good. Marginal value: compared to existing tools, CCCC provides cross-runtime coordination and persistent state, unique value. Cost-benefit: simple installation, no extra infrastructure, but learning curve, so scores 2.
Evidence shows: README claims (e.g., support for 17 runtimes) have corresponding documentation, but no independent verification. Cross-source corroboration: CI config and test scripts provide partial verification, but no external verification, so scores 1. Fact-inference separation: README clearly distinguishes stable and experimental features, so scores 2.
- Web UI is high-privilege; must set Admin Access Token before exposing publicly to prevent unauthorized access.
- Group Bridge read/full permissions should be granted cautiously, only to trusted remote groups.
- Rust implementation is experimental; use Python stable version for production.
- Dependency security lacks vulnerability scanning evidence; recommend checking dependencies for vulnerabilities.
What does this agent do, and when should you use it?
CCCC is a local-first coordination system for multi-agent software development, centered on a daemon that manages groups, actors, and message state. Its append-only JSONL ledger records messages and events, while a Web UI, CLI, MCP surface, and IM bridges use the same control plane. A group can combine 16 first-class runtimes, including Claude Code, Codex CLI, and ChatGPT Web, with a custom runtime path for other commands. It fits teams that need long-running coding work, auditable handoffs, and delivery/read/reply tracking, with runtime state stored in CCCC_HOME rather than the repository.
Run cccc attach . to bind a directory as the work scope, cccc setup to configure runtimes, and cccc actor add <id> --runtime <runtime> to create a foreman and peers. cccc group start launches actors; the daemon delivers cccc send and cccc tracked-send messages and appends messages and events to ledger.jsonl. Actors use the MCP collaboration core for inboxes, messaging, files, context, tasks, and state, while on-demand capabilities include cccc_repo, cccc_terminal, cccc_actor, cccc_automation, and remote-group access tools. The Web UI exposes chat, delivery/read state, embedded terminals, automation rules, and settings; bridges for Telegram, Slack, Discord, Feishu/Lark, DingTalk, WeCom, and Weixin expose coordination and status operations through IM.
- An engineering lead coordinating Claude Code and Codex CLI on one repository, with explicit read receipts and validation handoffs.
- A developer or team supervising long-running coding work from a phone through Telegram, Slack, or another supported IM bridge.
- A distributed team running CCCC groups on Windows, WSL, Macs, servers, or trusted teammate machines and using Group Bridge for explicit cross-group collaboration.
- An engineering organization that needs replayable message and event records to audit whether delegated work was delivered, read, acknowledged, and answered.
- A user who wants to bind a ChatGPT Web conversation as a development actor with browser delivery and a remote MCP connector.
What are this agent's strengths and limitations?
- A single-writer daemon and append-only ledger.jsonl provide replayable, auditable message and event state.
- Its coordination model includes routing, read cursors, attention acknowledgments, and reply-required obligations rather than only terminal prompting.
- Web UI, CLI, MCP, and IM bridges share daemon-owned state, supporting both local terminal control and phone-based operations.
- It supports 16 first-class runtimes plus custom, allowing combinations such as Claude Code, Codex CLI, and ChatGPT Web in one group.
- The default Python distribution requires Python 3.11+; the native distribution requires Rust 1.88+, and both cccc executables must not coexist on PATH.
- The ChatGPT Web runtime depends on browser delivery, remote MCP, and a public HTTPS exposure; GPT-5.x Pro sessions cannot use this integration.
- Daemon IPC has no authentication and binds to localhost by default; non-local Web exposure requires an Admin Access Token and a network boundary.
- Group Bridge read and full access can let trusted remote groups inspect context or operate on a workspace, so permissions require careful administration.
How do you install or deploy this agent?
The default distribution requires Python 3.11+: pip install -U cccc-pair. The native Rust distribution requires Rust 1.88+: cargo install cccc --locked. Do not keep both cccc executables on PATH. Then run cccc; it starts the daemon and local Web UI by default. Open http://127.0.0.1:8848.
How do you use this agent?
From the target repository, run: cccc attach .; cccc setup; cccc actor add foreman --runtime claude; cccc actor add implementer --runtime codex; cccc group start. Send ordinary coordination with cccc send "Please inspect the repo and propose the first safe task." --to foreman. For a durable delegation, run cccc tracked-send "Please take the first concrete task and reply with validation evidence." --to implementer --title "First concrete task" --outcome "The change and validation evidence are reported". Run cccc doctor if an actor stays stopped. The ChatGPT Web runtime additionally requires exposing CCCC through a public HTTPS URL and configuring a remote MCP connector in one bound ChatGPT conversation.
How does this agent compare with similar options?
Compared with native agent teams such as Claude Code subagents/teams, CCCC adds cross-runtime groups, state that survives restarts, phone/IM operations, and an audit ledger. Compared with isolated parallel task runners, it emphasizes messaging, handoffs, acknowledgments, and daemon-owned coordination; the README positions pure DAG orchestration as a case for a dedicated orchestrator.