Release v1.1.0

castiel - v1.1.0

Small new version update regarding the Cline provider. Now you can authenticate via OAuth besides API key and access your models by @ISO53 in #21 and #22

Full Changelog: v1.0.0...v1.1.0

Release v1.0.0

castiel - v1.0.0

This is the first stable release of castiel. After five tagged pre-releases (v0.0.0 → v0.7.6), the engagement workflow, the document model, and the tool-calling loop have hardened into something reliable enough to run real assessments against.

Highlights

• AI operator with a full tool-calling loop - shell execution, managed background processes, file read/write/edit, workspace search, web search and fetch, and delegated sub-agents, guided by a system prompt covering phase confinement, noise discipline, and rules of engagement.
• Human stays in charge - phase transitions happen only through the UI, scope boundaries are enforced from engagement.json, and the operator asks a structured question instead of guessing when it needs a decision.
• Five-phase engagement model - planning, recon, scanning, exploitation, wrap-up. The operator works inside the active phase; you decide when to advance.
• Structured engagement documents - findings live in plain JSON (network.json, web.json, vulnerabilities.json, evidence.json) with CVSS scoring, provenance stamps, evidence links, and user-owned note fields the operator never overwrites.
• Live document views - an interactive network topology graph, a ports & services matrix, a web-app map with entry points and tech stack, a CVSS speedometer with a filterable findings table, and an evidence table with inline previews, image lightbox, and masked secrets. Views refresh the moment the operator writes to disk.
• Workstation-style interface - streaming chat with reasoning traces and tool-call detail, a file tree toggling raw files and rendered views, live background-process and sub-agent docks, and an MCP panel with per-server enable/disable.
• Model freedom - works with llama.cpp, Ollama, OpenRouter, and Cline. Reasoning effort adapts per model, chat history is compacted automatically, and malformed model output is repaired instead of crashing the turn.
• Extensible via MCP - servers connect live from mcp.json without a restart and stand next to the built-in tools.

Changes since v0.7.6

• Evidence: table view with artifact detail dialog, image/text preview, image lightbox, secret masking, and a simplified evidence document schema - @ISO53 in #19
• Thinking: dynamic reasoning levels per model - @ISO53 in #18
• Tasks: removed the tasks document and kanban board entirely - @ISO53 in #17
• Tool call: handle truncated model output and malformed tool calls; reduced bash tool timeout defaults - @ISO53 in #16
• CI: upgraded GitHub Actions to v5 - @ISO53 in #15

Full Changelog: v0.7.6...v1.0.0

Release v0.7.6

castiel - v0.7.6

This PR fixes five important bugs across castiel, including tool-call and failure logging, MCP ping timeouts, chat history navigation, debug mode in development, and live runtime updates for background processes and sub-agents. It also includes dependency updates and several small runtime and UI improvements.

• Bug fix/chat by @ISO53 in #14

Full Changelog: v0.7.1...v0.7.6

Release v0.7.1

castiel - v0.7.1

Minor new version with website and release packaging updates.

• Rename artifacts to version-less names and disable macOS job by @ISO53 in #13

Full Changelog: 0.7.0...v0.7.1

Release v0.7.0

castiel - v0.7.0

This release focuses on strengthening castiel’s agent architecture and improving its runtime experience. It introduces sub-agents and orchestration, adds CLI configuration options and more robust optional logging, and improves workspace isolation by restricting AI access to internal chat data. The release also introduces automated builds for OS-specific executables, moving beyond distributing castiel as a single .jar file.

• Removed AI agents access to .chats directory in some tools by @ISO53 in #8
• Added sub-agents/orchestrator agent mechanism to the app by @ISO53 in #9
• Added couple basic CLI arguments to the app by @ISO53 in #10
• Added more robust logging mechanism that can be enabled by CLI args by @ISO53 in #11
• Added scripts and Github Actions workflows to create OS specific executables instead of single .jar files by @ISO53 in #12

Full Changelog: v0.3.0...0.7.0

Release v0.3.0

castiel - v0.3.0

This release brings several improvements to castiel, including support for Cline as an LLM provider, a new project website, and configurable MCP server toggling. It also includes various chat fixes, improved text selection styling, synchronized project versions, and a new bottom bar for a more complete interface.

• Added Cline to LLM providers by @ISO53 in #1
• Created website for the app by @ISO53 in #2
• Synced pom.xml versions with package.json by @ISO53 in #3
• Fix text selection with better coloring in website by @ISO53 in #4
• Added new feature that allows toggling MCP servers by @ISO53 in #5
• Various bug fixes on chat functionality by @ISO53 in #6
• Added bottom bar by @ISO53 in #7

Full Changelog: v0.0.0...v0.3.0

Release v0.0.0

castiel - v0.0.0

castiel is an AI-assisted penetration testing harness. It pairs a human engagement lead with an AI operator that handles the technical workload. Probing targets, reasoning about findings, and keeping a structured record of the entire engagement, while the human keeps full control over scope, phase, and authorization.

This is the first public release of castiel.

• AI operator with a full tool-calling loop: run shell commands, supervise background processes, read/write/edit files, search the web, and search the workspace.
• Currently supports llama.cpp, Ollama and OpenRouter for LLM providers.
• Five-phase engagement model (planning, recon, scanning, exploitation, wrap-up) that stays inside the active phase; the human triggers every phase transition.
• Findings recorded in plain JSON documents with CVSS scoring, evidence, and full provenance, so the engagement builds a clean, reviewable record.
• Extensible through MCP servers, picked up live without a restart, alongside the built-in tools.
• Workstation-style interface: streaming chat with reasoning, tool-call detail, cited sources, code highlighting; a file tree toggling raw files and rendered views; a live background-process panel; and a dedicated view per document. Network topology, web-app surface, and a CVSS gauge with a drillable findings table.