Repository navigation
Releases
Vibe Coding Hub v1.2.0
A big release focused on flows, external agent integration, and a much more polished first-run and day-to-day UI.
Highlights
- Agent Client Protocol (ACP) support, with external agents usable as characters
- Remote runners for executing work off your machine
- Headless flow runs in CI, with exportable run reports and an audit trail
- First-run wizard with automatic local provider detection, plus a sample project and a guided "Fix this test" first task
- A real interactive terminal and a new labelled activity bar
Flows
- Project-level flows
- Flow template gallery
- Per-step model routing
- Budget cap for flows
- Run history and replay from a step
- Approve with comments at gate steps
- Run flows headless in CI
- Export a flow run
- Markdown rendering in flow step output and background task output
Agents, Skills & Integrations
- Agent Client Protocol (ACP) support
- External agents as characters
- Remote runners
- MCP server manager
- Git hosting integration
- Skill and character packs
- Import skills and rules
- Audit trail and exportable run reports
- Privacy controls
Chat & Editing Experience
- One chronological chat timeline with smart auto-scroll
- Diff-first approval dialog with "allow for this session"
- Shift+Enter inserts a new line in the chat input
- Command palette: focus on open and better ranking
- Resume the last session on start; session list
- Requirements clarifier now runs only for greenfield / app-building requests
- Advanced options that already existed are now exposed in the UI
Onboarding & Updates
- First-run wizard with local provider auto-detection
- Sample project and a guided "Fix this test" first task
- In-app update check and one-click update
- Up-to-date model defaults and automatic model discovery
UI Polish
- Emoji icon toolbar replaced by a labelled activity bar
- Theme token pass for better contrast and readability
- Layout fixes: overlaps, clipping, alignment
- Local times shown everywhere
- Fixed tool cards and the terminal header
Bug Fixes
- Fixed Linux startup crash
- Fixed Session Dashboard crash on "Create session"
- Global exception handling: one bad view no longer takes down the app
- Stream parser now honours
finish_reasonarriving together with tool-call deltas - Logs now go to the per-user app-data folder instead of the project
- Tool names in presets and characters are now validated
Testing, CI & Tooling
- Benchmark harness
- Automated screenshot generation
- Scripted fake provider as a shared test fixture
- Hermetic tests throughout (WebFetch tests no longer depend on DNS)
- CI pipeline fixes, and stabilised macOS tests for remote server dispose and PTY Ctrl-C
Full changelog: v1.1.0...v1.2.0
What's Changed
- fix: linux startup crash by @Eisenmann in #6
- fix: global exception handling: never lose the app to one bad view by @Eisenmann in #7
- fix: Session Dashboard crash on "Create session" by @Eisenmann in #8
- fix: Make all tests hermetic by @Eisenmann in #9
- fix: Validate tool names in presets and characters by @Eisenmann in #10
- feat: Make WebFetch tests independent of DNS by @Eisenmann in #11
- fix: Write logs to the per-user app-data folder, not the project by @Eisenmann in #12
- fix: Stream parser: honour
finish_reasonthat arrives with tool-cal… by @Eisenmann in #13 - feat: Run the requirements clarifier only for greenfield / app-buildi… by @Eisenmann in #14
- fix: tool cards and the terminal header by @Eisenmann in #15
- feat: One chronological chat timeline with smart auto-scroll by @Eisenmann in #16
- feat: Shift+Enter inserts a new line in the chat input by @Eisenmann in #17
- feat: Diff-first approval dialog with "allow for this session" by @Eisenmann in #18
- feat: Command palette: focus on open and better ranking by @Eisenmann in #19
- fix: Show local times everywhere by @Eisenmann in #20
- feat: Render Markdown in flow step output and background task output by @Eisenmann in #21
- feat: Resume the last session on start; session list by @Eisenmann in #22
- fix: Up-to-date model defaults and automatic model discovery by @Eisenmann in #23
- feat: First-run wizard with local provider auto-detection by @Eisenmann in #24
- feat: Sample project and a guided "Fix this test" first task by @Eisenmann in #25
- feat: In-app update check and one-click update by @Eisenmann in #26
- feat: Replace the emoji icon toolbar with a labelled activity bar by @Eisenmann in #27
- feat: Theme token pass: contrast and readability by @Eisenmann in #28
- feat: Layout fixes: overlaps, clipping, alignment by @Eisenmann in #29
- feat: A real interactive terminal by @Eisenmann in #30
- feat: expose the advanced options that already exist by @Eisenmann in #31
- feat: project-level flows by @Eisenmann in #32
- feat: flow template gallery by @Eisenmann in #33
- feat: per-step model routing by @Eisenmann in #34
- feat: budget cap for flows by @Eisenmann in #35
- feat: Run history and replay from a step by @Eisenmann in #36
- feat: Approve with comments at gate steps by @Eisenmann in #37
- feat: Run flows headless in CI by @Eisenmann in #38
- feat: export a flow run by @Eisenmann in #39
- feat: skill and character packs by @Eisenmann in #40
- feat: import skills and rules by @Eisenmann in #41
- feat: mcp server manager by @Eisenmann in #42
- feat: git hosting integration by @Eisenmann in #43
- feat: Agent Client Protocol (ACP) support by @Eisenmann in #44
- feat: external agents as characters by @Eisenmann in #45
- feat: Audit trail and exportable run reports by @Eisenmann in #46
- feat: remote runners by @Eisenmann in #47
- feat: privacy by @Eisenmann in #48
- feat: scripted fake provider as a shared test fixture by @Eisenmann in #49
- feat: screenshot generation by @Eisenmann in #50
- feat: benchmark harness by @Eisenmann in #51
- fix: ci pipeline execution by @Eisenmann in #52
Full Changelog: v1.1.0...v1.2.0
Vibe Coding Hub v1.1.0
Vibe Coding Hub v1.1.0
This release adds characters with their own skills and team flows: characters working together on a task, drawn on a canvas and run in parallel where the work allows.
Highlights
- Skills registry. Write skills as plain markdown files, then add, edit and delete them from the app or the CLI.
- Characters with unique skillsets. Give each character exactly the skills it needs. One developer can know clean architecture and C#, another ML and Python.
- Team flows. Connect characters with arrows into a flow. A character starts when the characters pointing into it are done, and independent characters run in parallel.
- Flows panel. Draw, validate and run flows in the desktop app, with live status for every step.
Skills and characters
Skills registry
- A skill is a markdown file with YAML frontmatter:
<dir>/<name>/SKILL.md. - Skills live in
~/.aiagent/skills(global) or.aiagent/skills(per project). A project skill overrides a global one with the same name. - Renaming or deleting a skill updates the characters that use it, and the doctor reports dangling references.
- Agents load skills on demand with the
use_skilltool. A character can only use the skills it has. - Pinned skills are always included in the character's context.
Characters
- A character is a markdown file in
characters/<id>.md, global or per project, with a persona and a list of skills. - Add skill button in the character profile: pick a skill from the library or create a new one on the spot.
- Templates: a character can extend a template (
extends) and add or remove skills (remove-skills), so a "software developer" base can become a C# architect or an ML engineer. spawn_subagentcan start a sub-agent as a specific character.- CLI:
aiagent characters add-skilland the other skill and character commands.
Note: the earlier "skill sets" idea was dropped in favor of per-character skills. An old skillsets: field in a character file is ignored and produces a warning.
Team flows
What a flow is
A flow is a pipeline whose stages are characters connected by arrows.
design (Alex) ──► backend (Dana) ──┐
└─► model (Mia) ─────┴─► review (Quinn) ──approved──► release notes (Sam)
▲ │
└──── rejected, ≤2× ──┘
- Parallel by default. Steps whose inputs are ready run at the same time, each in its own git worktree. The limit is
maxParallel(default 3). - Handoff. A step receives the task plus the final answers of the steps connected into it.
- Conditions. A step can declare outcomes such as
approvedorrejected. Arrows can follow a specific outcome. - Loops. An arrow back to an earlier step re-runs it with the feedback, up to a limit. When the limit is reached, a
loop-exhaustedarrow can route elsewhere. A cap of 25 step runs per flow prevents runaway runs. - Failures only skip the steps that depend on the failed one. Other branches finish.
- Confirmation prompts and cancellation work per step.
- Flows are stored in
pipelines/*.json. The new fields are optional, so existing pipelines, includingfull-sdlcandquick-fix, behave as before.
Flows panel (desktop)
Open it with the 🔀 button or from the command palette ("Flows", "New Flow").
- Draw: drag characters from the palette, move steps, drag from a step's right handle to connect. Zoom, pan, undo and redo are supported.
- Edit: set each step's character, prompt, outcomes, confirmation and on/off state, and each arrow's condition and loop limit.
- Check: problems are outlined in red with a list below, and the header shows which steps run in parallel.
- Auto-layout arranges steps by execution order.
- Run: each step shows waiting, running, done, failed or skipped. Arrows light up when taken. Click a step to see its live output and answer approvals on the step. Stop cancels the run.
CLI
aiagent pipeline runshows per-step progress and accepts--max-paralleland--cast.aiagent pipeline validate <name>lists problems and exits with code 1 on errors.aiagent pipeline show <name> [--mermaid]lists the execution waves and loops, or exports a Mermaid diagram.
Examples
examples/pipelines/feature-team.json: Alex designs, Dana (backend) and Mia (model) work in parallel, Quinn reviews and loops on rejection, Sam writes the release notes.- The README has a new section, Flows: characters working together.
Worktree improvements
- Parallel agents now see uncommitted work from the main checkout, such as files written by an earlier step. Their results are applied back as uncommitted changes, so nothing is committed to your branch.
- Each loop round gets its own branch.
- A merge conflict fails that step and blocks what depends on it. The branch and worktree are kept so you can resolve it by hand.
Fixes
- The chat's Run pipeline button works again.
- The chat shows events from flow steps, and its pipeline picker refreshes when a flow is saved.
- Several cross-platform fixes in tests and git handling: macOS symlinked temp folders, Windows line endings, and timing-dependent tests.
Upgrade notes
- No action is needed for existing pipelines or characters.
- Saving a linear pipeline from the Flows panel converts it into a flow. The panel warns you first.
- Remove any leftover
skillsets:fields from character files.
What's Changed
- Feature/skills registry by @Eisenmann in #4
- feat: team flows by @Eisenmann in #5
Full Changelog: v1.0.1...v1.1.0
v1.0.1
v1.0.1
First release with downloadable builds. The application itself is unchanged from v1.0.0.
Added
- CI pipeline: every push to
mainand every pull request is built and tested on Windows, Linux and macOS. - Release pipeline: pushing a
v*tag builds self-contained, single-file binaries and attaches them to the GitHub Release:- Windows x64 (
.zip) - Linux x64 (
.tar.gz) - macOS Intel and Apple Silicon (
.tar.gz) - each as the desktop app (Vibe Coding Hub) and the CLI.
- Windows x64 (
Included from v1.0.0
- Parallel tasks across multiple projects.
- Explorer that updates automatically as files change, with git status badges.
- Source Control panel (Ctrl+Shift+G): view changes, stage, commit, discard.
- GitHub-style side-by-side diff viewer with word-level highlights and per-hunk revert.
- Live activity bar in chat showing the current action and elapsed time.
- Fix: Send now works when the agent asks a clarifying question.
Notes
- Binaries are not yet code-signed or notarized. Windows SmartScreen and macOS Gatekeeper may show a warning on first launch.
What's Changed
- ci: add build and release workflows by @Eisenmann in #3
Full Changelog: v1.0.0...v1.0.1
v1.0.0
Vibe Coding Hub v1.0.0
The first stable release of Vibe Coding Hub: an AI coding agent with a cross-platform desktop app and a CLI, built on .NET 10. It works with OpenAI, Anthropic, Ollama and any OpenAI-compatible API (including LM Studio), so you can use hosted or fully local models.
Highlights
- Desktop app (Avalonia) with a chat view, code editor with diffs, file preview, command palette and a multi-agent session dashboard.
- CLI for interactive chat, one-shot prompts and SDLC pipelines.
- Multi-agent sessions: run agents in parallel, let them message each other, and see which agent produced each change.
- Safe by default: permission modes, checkpoints/undo, optional Docker sandbox for shell commands.
Features
Agents and workflows
- Autonomous multi-file agent that shows a visible plan before and while it works.
- Subagents for delegating scoped tasks (
spawn_subagent). - Multi-agent coordination: parallel agent groups, an inter-agent mailbox (
send_message, direct or broadcast), and per-agent diff attribution. - Git worktree isolation for parallel agents: each works on its own branch and is merged back one at a time. Conflicts keep the branch for manual resolution.
- SDLC pipelines (analyze, implement, review, test, deploy) with built-in roles (
planner,implementer,reviewer,tester,deployer). Pipelines and roles are plain JSON under~/.aiagent/pipelinesand~/.aiagent/presets. - Conversational requirements clarification (
ask_user) before the agent starts building.
Code intelligence
- Whole-codebase semantic Q&A over an indexed workspace.
- Language Server (LSP) based code intelligence.
- Predictive next-edit autocomplete.
- Click-to-edit visual layer on generated UI, and freehand visual annotation.
- In-chat branch / PR workflow and source control view.
Context, memory and customization
- Session persistence and resume, plus session export/import.
- Checkpoints, undo,
/rewindand/forkto go back to or branch from any message. - Context window management with automatic trimming and compaction.
- Project memory (
AGENTS.md) and auto memory for learned preferences. - Skills, hooks, MCP connections and bundled backend primitives.
- Model configuration with a reasoning-effort selector (Default/Off/Low/Medium/High).
Tools
read_file, write_file, edit_file, multi_edit, apply_patch, list_directory, glob, grep, execute_command, git, web_fetch, todo_write, ask_user, run_diagnostics, verify_changes (build/test/lint, stops after 5 failed attempts in a row), and browser_check (headless Chrome/Edge: console, errors, failed requests, screenshots).
Cost, safety and isolation
- Prompt caching for Anthropic, with live token, session cost and cache-hit display. Optional spend cap via
AgentOptions.MaxBudgetUsd; prices can be overridden in~/.aiagent/pricing.json. - Permission modes, read-only mode, allowed paths and dangerous command detection.
- Optional sandboxed shell with
AIAGENT_SANDBOX=docker(network off, capabilities dropped, workspace mounted). If Docker is unavailable the command is refused, never run on the host.
Downloads
Self-contained single-file builds, no .NET install needed:
| Platform | Desktop app | CLI |
|---|---|---|
| Windows x64 | vibe-coding-hub-1.0.0-win-x64.zip |
vibe-coding-hub-cli-1.0.0-win-x64.zip |
| Linux x64 | vibe-coding-hub-1.0.0-linux-x64.tar.gz |
vibe-coding-hub-cli-1.0.0-linux-x64.tar.gz |
| macOS Intel | vibe-coding-hub-1.0.0-osx-x64.tar.gz |
vibe-coding-hub-cli-1.0.0-osx-x64.tar.gz |
| macOS Apple Silicon | vibe-coding-hub-1.0.0-osx-arm64.tar.gz |
vibe-coding-hub-cli-1.0.0-osx-arm64.tar.gz |
Getting started
# Set a provider key (or use AIAGENT_OPENAI_API_KEY / AIAGENT_ANTHROPIC_API_KEY)
aiagent config set-key anthropic sk-ant-...
# Interactive chat, or a single prompt
aiagent chat --provider anthropic
aiagent run "Add unit tests for the UserService class"Known limitations
- Worktree-isolated agents start from
HEAD, so uncommitted changes in your checkout are not visible to them. - Built-in model prices are estimates; cost is shown only when a model is set explicitly. Unknown models (e.g. Ollama) are treated as free.
- Release binaries are not code-signed, so Windows SmartScreen and macOS Gatekeeper may warn on first launch.
License
MIT. Full commit history: https://github.com/AIHelpers/vibe-coding-hub/commits/v1.0.0