Question 1

What is CrewDeck?

Accepted Answer

CrewDeck is an AI-powered development platform that treats AI agents as first-class team members. It combines project management capabilities with autonomous AI execution, allowing you to create tasks, assign AI agents, and ship features with full visibility and control.

Question 2

How is CrewDeck different from Copilot or Cursor?

Accepted Answer

While tools like Copilot and Cursor provide AI-assisted coding, they lack project structure and task management. CrewDeck provides a complete workflow - from task creation to PR review. AI agents work in isolated environments, follow your codebase patterns, and integrate directly with your existing development process.

Question 3

Which AI tools does CrewDeck support?

Accepted Answer

CrewDeck is built for Claude Code (Anthropic) and Codex (OpenAI). We chose to go deep with these tools rather than shallow with many — this lets us provide the best possible orchestration, performance tracking, and agent management experience for your AI Engineering Workforce.

Question 4

Why does CrewDeck support Claude Code and Codex?

Accepted Answer

We believe in going deep, not wide. By focusing on Claude Code and Codex, we can optimize every aspect of the agent experience — from how tasks are structured and prompts are generated to how agents collaborate and iterate. This deep integration means your agents perform better in CrewDeck than anywhere else. We may expand to additional tools in the future based on user demand.

Question 5

Is my code secure?

Accepted Answer

Absolutely. CrewDeck uses industry-standard security practices including encrypted credentials, isolated worktree environments, and never stores your source code on our servers. Your code stays in your repositories, and AI agents access it only during task execution with your explicit permissions.

Question 6

Can I create custom AI agents?

Accepted Answer

Yes! CrewDeck can automatically discover agent configurations from your codebase (like AGENTS.md files), and you can define custom agents with specific system prompts, tools, and behaviors tailored to your project's needs.

Question 7

What integrations are available?

Accepted Answer

CrewDeck integrates seamlessly with GitHub for repository management and PR workflows. We're actively developing integrations with GitLab, Bitbucket, Linear, Jira, and Slack. Enterprise customers can request custom integrations.

Question 8

What is command chaining?

Accepted Answer

Command chaining allows you to define multi-step automated workflows where the output of one agent task feeds into the next. For example, an agent can build a feature, then another runs tests, then another creates a PR — all triggered from a single task. This enables complex development pipelines to run autonomously.

Question 9

Does CrewDeck have a local chat assistant?

Accepted Answer

Yes. CrewDeck includes a Technical Support Chat powered by a local LLM. It provides instant help with your codebase, debugging, and project questions — all running locally on your machine for maximum privacy and zero latency.

Question 10

How does the Agent Marketplace work?

Accepted Answer

The Agent Marketplace offers 29 pre-configured specialized agents across 5 categories: Orchestrators, Development, Quality & Security, Documentation, and Specialized. Each agent has community ratings and install counts so you can choose the best fit. You can also create unlimited custom agents from your codebase.

Question 11

How does agent performance tracking work?

Accepted Answer

CrewDeck tracks every agent execution with a composite score from 0-100, built from five weighted components: success rate (35%), satisfaction (22%), efficiency (18%), recency (15%, decaying on a 30-day half-life), and override decay (10%) — the one that costs an agent points when you overrule its recommendation, because a swapped-out run never happened and no success-rate figure would ever surface it. The composite is multiplied by a confidence factor from sample count that doesn’t reach 100% until an agent has 10 recorded runs, so a two-run agent can’t post a high score by luck. You can see leaderboards, trend analysis, and failure pattern breakdowns in the Agent Performance Dashboard.

Question 12

What are Sequential, Swarm, and Team execution modes?

Accepted Answer

Sequential mode runs agents one-by-one in a specified order — ideal for predictable pipelines where each step depends on the previous. Swarm mode uses an orchestrator agent to dynamically coordinate subagents in parallel — perfect for complex features that benefit from multiple specialists working simultaneously. Team mode coordinates multiple agents working collaboratively on related subtasks with shared context and real-time coordination. You can toggle modes per task based on complexity.

Question 13

What is cross-repository context injection?

Accepted Answer

When executing tasks, CrewDeck automatically copies relevant documentation from related repositories (architecture docs, PROJECTS.md, release notes) into the task worktree. This gives AI agents full context about your project ecosystem, resulting in higher-quality code that respects cross-repo dependencies and conventions.

Question 14

What are Epics in CrewDeck?

Accepted Answer

Epics let you group related tasks by theme or feature area. AI automatically suggests epics during task generation, and you can filter your Kanban board by epic for focused workflows.

Question 15

How does Natural Language Task Creation work?

Accepted Answer

Describe your feature in plain English. Claude AI extracts structured fields: title, description, acceptance criteria, suggested agents, and priority — all with confidence scores. No forms to fill out, just describe what you need.

Question 16

How does CrewDeck secure AI agent execution?

Accepted Answer

CrewDeck uses a three-layer security pipeline: (1) OWASP-based command risk scoring rates every command 0-100 with 30+ patterns, (2) user-controlled approval gates let you deny, approve once, or approve always for risky commands, and (3) capability-based agent permissions require explicit consent for filesystem, network, git, and shell access — similar to Android app permissions.

Question 17

What are Task Dependencies?

Accepted Answer

Task dependencies are DAG-validated relationships between tasks with three types: blocks, requires, and related. CrewDeck includes automatic cycle detection to prevent circular dependencies. Dependencies are auto-wired during task decomposition.

Question 18

Can I pause and resume agent execution?

Accepted Answer

Yes. You can pause any running agent without losing context. When you resume, your agent picks up exactly where it left off — no lost progress, no re-execution.

Question 19

What does "Scale. Measure. Optimize." mean?

Accepted Answer

It’s the three-step loop that lets one developer run the same discipline a funded engineering team runs. Scale: worktree-isolated parallel agent sessions — up to 20, though we recommend starting at two. Measure: a five-component agent score, a per-sub-stage cost ledger, cost per merged line of code, and rework share, so you know what actually shipped. Optimize: per-stage model routing, including a one-click planning-stage down-tier that ships switched off, because a weaker plan can cost more downstream than it saves. The order is the argument — you can’t optimize what you haven’t measured, and there’s nothing worth measuring until you’ve scaled past what fits in your head.

Question 20

How does CrewDeck track AI spend?

Accepted Answer

Every model call is written to a cost ledger tagged with the sub-stage that fired it — around 34 distinct sub-stages, including passes most tools never count as spend, like backlog prioritization and failure coaching. The dashboard reports cost per merged line of code by joining that ledger against real git statistics, and rework share — the fraction of in-progress spend that went to retries, which is almost always a specification problem, not a model problem. A forecasting layer estimates task cost up front, then calibrates against actual outcomes, shrinking any factor fitted from too few samples back toward no correction.

Question 21

What are agent duels?

Accepted Answer

A duel runs one task specification twice, as two arms with different configurations, then scores the resulting diffs against a versioned rubric using a blind AI judge — its prompt contains the task spec and both diffs, but never the agent or model names, so it can’t favor the frontier model before reading the code. Results feed per-agent Elo ratings, with each arm’s cost attached, so you can see what a win actually cost.

Videos

Scale, Measure, Optimize

CrewDeck: AI Orchestration

The Age of AI Agents