Craft Agents: Document-centric multi-provider agent desktop platform
Craft Agents is a document-centric desktop agent platform that connects multiple LLMs, MCPs and REST APIs to enable shareable sessions and seamless external-service integration for teams embedding agents into document workflows.
GitHub lukilabs/craft-agents-oss Updated 2026-04-18 Branch main Stars 5.8K Forks 771
Electron desktop Multi-LLM connections MCP integration Document-centric workflow Open-source (Apache-2.0)

💡 Deep Analysis

2
How should permissions and automation policies be configured in Craft Agents to balance efficiency and security?

Core Analysis

Core Question: With Craft Agents’ three-tier permission model, how to design policies that allow efficient automation while controlling accidental operations and data leakage?

Technical Analysis

  • Permission granularity: Use workspace and skill-level isolation to mark write-capable or external-change skills as high-risk.
  • Staged escalation: Move from Explore (read/simulate) → Ask to Edit (requires human confirmation) → Auto (fully automatic) to reduce blast radius.
  • Credentials & audit: Combine a vault for minimal-permission credentials with detailed API call and session audit logging for traceability.

Practical Recommendations

  1. Default security template: New workspaces default to Ask to Edit or Explore until compliance checks pass.
  2. Force human approval for write ops: Skills that can modify external systems (send emails, update DBs, create tasks) should require confirmation.
  3. Version skills & backup sessions: Enable versioning and quick rollback for skill changes.
  4. Monitoring & alerts: Trigger alerts and auto-pause automation when agents perform sensitive ops or exceed thresholds.

Caution

Important: Do not set Auto as the default; always use Ask to Edit for public or high-sensitivity scenarios.

Summary: By combining workspace/skill-level permissions, staged escalation, credential minimization and audit/rollback, you can balance automation efficiency with acceptable security risk.

90.0%
What is the day-to-day user experience of Craft Agents? What is the learning curve and common pitfalls?

Core Analysis

Core Question: Evaluate the real-world UX for engineers and non-engineers using Craft Agents, including the learning curve and common pitfalls.

Technical Analysis

  • Strengths: Natural-language source addition, session/document-centric interaction, streaming responses, and visualized tool outputs reduce cognitive load. Best experience is with public APIs, well-formed OpenAPI specs, and simple local MCP subprocesses.
  • Pain points & pitfalls:
  • Private networks & jumpboxes: Auto-configuration often needs manual network and credential setup.
  • Permission risks: Default or Auto modes can trigger unintended write actions.
  • File & OpenAPI parsing errors: Complex documents may be misconverted and require manual checks.
  • External model limits: Quotas, rate limits, and cost impacts UX directly.

Practical Advice

  1. Gradually escalate permissions: Explore → Ask to Edit → Auto.
  2. Sandbox private sources first: Validate each source/skill in an isolated workspace.
  3. Use a secrets vault & minimal-permission keys, rotate them regularly.
  4. Keep human review for critical skills, especially those performing writes.

Caution

Important: Do not set a workspace to Auto without verification; always manually verify complex document parsing.

Summary: Craft Agents substantially improves day-to-day agent workflows but achieving reliable production behavior requires accompanying ops and security practices.

88.0%

✨ Highlights

  • Document-centric seamless multi-provider agent support
  • Streaming responses with multi-session management
  • Repository metadata shows inconsistent activity indicators
  • No releases or contributor records in provided metadata—adoption caution advised

🔧 Engineering

  • Document-native agent UI enabling intuitive multitasking and shareable sessions
  • Supports multiple LLMs and MCP/REST data sources with highly automated import and connection flows

⚠️ Risks

  • Provided data shows no contributors or commits; may indicate scraping omissions or weak actual maintenance
  • Interactions with external APIs and MCPs involve credentials and permissions—deployment and operations require security and compliance care

👥 For who?

  • Product and design teams aiming to embed agents into document workflows to improve collaboration
  • Enterprises and research teams needing private/remote long-running sessions and customizable toolchains