Deckent is building toward a fully autonomous AI development assistant — in the same category as OpenClaw, Microsoft Copilot Cowork, and Devin. The current phase is AI agent orchestration: a multi-agent CLI that plans, executes, and evaluates development tasks in parallel sprints. This is not the destination — it is the foundation.
The long-term goal: Deckent becomes an always-on, self-improving development teammate. It understands your codebase, learns from every sprint, plans proactively, and operates with minimal human input. Open source, self-hosted, provider-agnostic — the anti-Devin.
Solo AI assistants are inherently limited: one context window, one task, one perspective. Deckent breaks through this ceiling with its Brain-Worker-Auditor architecture. Brain sets the strategy, Workers execute in parallel, Auditor guarantees quality. After every sprint, learnings persist to memory — the system makes better decisions with each iteration.
Where we are now: AI orchestration CLI — sprint-based multi-agent execution with 3 spawn backends (tmux, subprocess, Docker), 3 AI providers, 16 agents, 21 skills.
Where we are going: Autonomous AI assistant — heartbeat daemon, proactive task execution, channel integrations (Slack, GitHub), codebase semantic understanding, always-on gateway. Think OpenClaw's architecture + Deckent's multi-agent discipline.
| Segment | Profile | Deckent Value |
|---|---|---|
| Individual developer | Indie dev, freelancer, solo founder | Multi-agent power for a one-person team — parallel output through sprints |
| Small team | 2-10 person startup or squad | Use AI workers as team members — automate repetitive tasks |
| Enterprise | Large-scale organization (future) | Controlled autonomous development — audit trail, scope enforcement, memory/learning |
| Tool | Category | Strength | Weakness | Deckent Position |
|---|---|---|---|---|
| OpenClaw | Autonomous AI assistant (343K+ stars) | Always-on daemon, 13K+ skills, 50+ channels | Single agent, no sprint lifecycle, no scope enforcement | Multi-agent orchestration + sprint discipline + learning |
| Copilot Cowork | Enterprise AI orchestrator | Multi-model critique layer, M365 integration | Closed source, $30+/user/month, no self-hosted | Open source, self-hosted, free, provider-agnostic |
| Devin | Autonomous software engineer | End-to-end autonomous, interactive planning | Single agent, closed source, $20-500/month | Multi-agent parallel, open source, free |
| Perplexity Computer | Multi-model AI agent | 19 models, days-long tasks, 400+ apps | $200-325/month, no self-hosted, no sprint planning | Self-hosted, 13 models, sprint-based structure |
| Claude Code (solo) | Single AI assistant | Strong single-task performance | Single context, no parallelism | Uses Claude Code as a worker, adds orchestration |
Deckent's unique position: The only open-source tool that combines multi-agent parallel execution + sprint lifecycle + scope enforcement + memory/learning + multi-provider support + self-hosted. Current phase: orchestration CLI. Next phase: autonomous assistant (OpenClaw/Cowork class).
TypeScript delivers type safety that enables confident refactoring across large codebases. ESM (ES Modules) aligns with the modern Node.js ecosystem and unlocks optimizations like tree-shaking. For a system that manages AI agents, type safety is critical — a malformed config or task structure can crash an entire sprint.
Depending on a single AI provider creates both cost and availability risk. Deckent's provider-agnostic architecture assigns different models to different tasks: opus for complex architectural decisions, haiku for simple documentation. A provider fallback chain ensures resilience against outages.
Three backends for different contexts: tmux (fastest, live terminal, Linux/macOS default), subprocess (Windows fallback, file-based tracking), Docker (container isolation, resource limits, CI/CD ready). Each worker runs in its own isolated environment regardless of backend.
MCP integrates Deckent with any MCP-compatible IDE or tool. With 21 tools and 8 resources, the entire sprint lifecycle is programmatically accessible. This makes Deckent not just a CLI, but a platform.
Workers run in isolated Docker containers with memory limits, non-root execution, and volume-mounted auth. Project filesystem access is controlled per-container. This is the foundation for enterprise deployment, CI/CD integration, and future Kubernetes scaling.
Core sprint lifecycle, multi-agent parallel execution, tmux/subprocess backends, MCP integration, multi-provider support (Claude + Codex + Gemini), ModelRegistry, agent/skill ecosystem, heartbeat daemon, human checkpoints, adaptive thresholds.
Docker container backend (live-verified Sprint 119-129, 10 e2e tests, configurable timeout), documentation consolidation (BETA-TRACKER, i18n generators, docs.json automation), ERRORS.md active logging, backend smoke testing (tmux + subprocess + Docker via MCP + CLI), dashboard backend badge, ADR-027 hybrid backend decision, version 0.4.0-beta.1. Sprint 130: MCP instructions accuracy fix (21 tools), decision-engine V1 @deprecated archive + ADR-028, real coverage measurement (89.33%).
VerhexIO/deckent open-source repository, CI/CD pipeline (GitHub Actions + Docker backend), npm publish, .detect-secrets, CONTRIBUTING guide, community onboarding.
The leap from orchestration CLI to autonomous AI assistant:
- Always-on gateway — daemon mode, SSE dashboard, remote control
- Channel integrations — Slack bot, GitHub Issues/PR automation, Linear/Jira sync
- Codebase semantic understanding — AST indexing, dependency graph, RAG-enhanced context
- Multi-sprint chaining — days-long autonomous task execution
- Critique layer — multi-model verification (writer + reviewer pattern)
- Browser/Computer Use — Claude Computer Use SDK integration
- Provider expansion — Grok, Llama, Mistral, DeepSeek (ModelRegistry ready)
This is where Deckent enters the OpenClaw/Cowork/Devin category — not as another single-agent tool, but as the only open-source multi-agent autonomous development platform.
- Open source — Deckent is free and open source. Community contributions are welcome.
- Transparency — Every sprint's plan, outcome, and learnings are on record. The
.brain/directory preserves the full decision history. - Quality — Auditor quality gates, GO/NO-GO evaluation, and mandatory testing ensure every sprint meets quality standards.
- Autonomous yet controlled — Deckent operates autonomously, but the user is always in control. Scope enforcement, audit trails, and memory budgets set clear boundaries.
- Continuous learning — MEMORY.md and PATTERNS.md are updated after every sprint. The system improves its decisions over time and avoids repeating the same mistakes.
- Orchestration first, autonomy next — Deckent starts as a sprint-based orchestrator and evolves toward full autonomy. Each phase builds on the previous — no shortcuts, no half measures.
| Metric | Value |
|---|---|
| Version | 1.0.0-beta.1 |
| Sprint | sprint-162a (T4 hot fix) |
| MCP Tools | 29 (verified) |
| MCP Resources | 8 |
| CLI Commands | 48 (verified) |
| Dashboard Pages | 6-7 |
| Agents | 18 (15 built-in + 3 custom) |
| Skills | 21 built-in |
| Providers | 3 (Claude, Codex, Gemini) |
| ADRs | 47 (44 accepted + ADR-047 NEW multi-lang) |
| Multi-language stacks | 6 baseline (TS/Py/Go/Rust/Java/C#) |
| Sprint | Status | Notable |
|---|---|---|
| sprint-161 | FAILED (mid-flight orchestration broke) | God-audit produced 356 findings, 8 bugs identified |
| sprint-162-failed-build-race | FAILED (build-after-spawn race) | dist mtime > sprint-state by 22s; Brain loaded old code |
| sprint-162a | T4 IMPLEMENTATION ✓ | 8 bug closed + multi-lang adapter + 4 ADR + 12-lang i18n + ARIA + security review |
| sprint-162-live | PARTIAL SUCCESS | 8/8 known bugs LIVE proven; Bug X discovered (dual-evaluator stale-state path race); org quota cascade at 15:36Z |
| Metric | Value |
|---|---|
| Sprint | sprint-162a |
| Status | PARTIAL — 8/8 bug closed + Bug X discovered |
| Total deliverables | 8 fix-spec + 14 source + 14 test + 4 ADR + 12-lang i18n + ARIA + security |
| Tests added | ~305 |
| Source LoC | +2703/-801 (69 files) |
| Beta GA blocker | Bug X cost cascade (Sprint 162B P0) |