Stop letting AI agents run unchecked in siloed sandboxes. Cordn8 establishes a hardened, append-only coordination channel directly in your repository (docs/agent-channel.md). Command multi-agent workflows with explicit autonomy dials, enforce grounding through a cryptographically anchored Context Spine, and run automated launch gates via the Maya validation engine. Built for teams standardizing on PR-driven engineering.
Plus a built-in local Next.js cockpit
(the Cordn8 Mantle Dashboard) — run
cordn8 dashboard
to monitor cost ledgers, latency metrics, and real-time agent trust scores directly on disk. No hosted service required.
or npx cordn8 start
to scaffold and run onboarding interactively
What it does
Cordn8 is the operational spine for repo-centered AI engineering. By formalizing coordination into an append-locked channel (docs/agent-channel.md), both human operators and autonomous agents collaborate using standardized git-visible handoffs. The Context Spine locks down a secure local Source Vault and Decision Ledger, ensuring agents are dynamically grounded against immutable context baselines. Autonomy is dialed fluidly from high-security MANUAL approvals to speed-optimized AUTOPILOT sandboxes. Before merging, the pluggable Maya Validation Engine executes a checklist of 22 key operator journeys and 10 structural gates to generate a launch-readiness scorecard—guaranteeing every AI-generated commit is safe, auditable, and production-ready.
Demo
demo.gif — coming soon
Recording in progress: cordn8 init → watch → do "build login" → channel tail.
How it's different
Cordn8 is built for teams that already standardise on PR-driven engineering and want multiple agents in the loop without moving coordination state out of the repo. Most of the listed alternatives are good at what they do — Cordn8 sits at a different point in the stack.
All coordination lives directly in your repository at docs/agent-channel.md. Track agent dispatches, approvals, and handoffs inline — fully git-visible and grep-able.
Onboarding is hardened against infinite validation loops. Automatically detects local runtimes like Ollama (brew, curl, or app) and falls back to a sandbox Tutorial Mode if credentials are absent.
Control agent velocity dynamically with MANUAL (requires human approvals), COPILOT (auto-approves low-risk steps), or AUTOPILOT (budget-capped autonomy).
Ground agent dispatches in a secure local Source Vault, locked Decision Ledgers, and frozen Mission Briefs. Guarantees agents always run against trusted context baselines.
Built-in validation decision engine. Runs 22 operator journeys and 10 structural gates (including the J13 cold-start sentinel) to verify project launch-readiness before merge.
Dispatch tasks uniformly to Anthropic Claude, OpenAI GPT, Groq, Ollama, Hermes, or your own custom-built CLI or API tools. Switch providers by changing a single config line.
A gorgeous local Next.js GUI launched with cordn8 dashboard. Real-time cost trackers, token latencies, and agent trust scoring. Works entirely on disk.
Concurrency-safe channel locks and domain allow-lists (allowing secure AI emails like `@anthropic.com`) guarantee agents commit and push changes safely under automated pipelines.
Quick start
Establish your repository's security posture, add grounding context, launch the local Next.js cockpit, and dispatch goals safely under explicit budgets.
# 1. Install the global Cordn8 CLI & local daemon npm install -g cordn8 # 2. Initialize the append-only channel & Context Spine in your repo cordn8 init # 3. Set your autonomy dial (MANUAL | COPILOT | AUTOPILOT) cordn8 mode set COPILOT # 4. Ground agent dispatches by adding trusted local source files cordn8 sources add "src/auth/" # 5. Spin up the watch daemon & open the local cockpit dashboard cordn8 watch & cordn8 dashboard # 6. Dispatch a goal with automated agent matching & budget cap cordn8 do "Refactor the session middleware" --budget 1.50 # 7. Run pluggable Maya release validation gates before merge cordn8 maya launch-check
Architecture
Cordn8 is architected from the ground up for strict repository sovereignty and zero network leakage. The core CLI operates as a local daemon, writing state securely to an append-locked markdown coordination channel (docs/agent-channel.md) featuring automatic schema validation, snapshots, and git-reconstructable history. Hardened telemetry—including step-by-step cost tracking, latency profiles, and trust scores—is persisted locally in an encrypted SQLite database. Provider integrations sit entirely behind a Universal Adapter Surface network boundary, giving you the choice to wire up air-gapped local runtimes (like Ollama and Hermes) or route securely to external clouds. The local Next.js Cordn8 Mantle Dashboard runs on disk, reading the exact same state as the CLI for a unified, completely off-grid monitoring experience.
channel
Append-only Markdown, concurrency locks, git-reconstructable
telemetry
Encrypted local SQLite — cost, latencies, trust scores, metrics
adapters
Air-gapped runtimes / Universal edge API boundaries — fully local option