VibeSpace is a local-first AI agent that runs on your Mac. Authorize a folder, describe the outcome in plain language, and it breaks the work down, reads code, runs commands, edits files, and delivers — every step visible, approvable, and recoverable.
0
Never uploads
100%
Runs locally
0
Web login optional
Your data, tasks, and keys stay on your Mac — and which provider you connect, what you pay, and where your content goes is entirely your call.
Claude, OpenAI, Gemini, DeepSeek, Kimi, Qwen, GLM, MiniMax, Doubao, and OpenRouter ship built-in, plus any OpenAI-compatible custom endpoint — paste an API key and you're in.
Model requests go straight from your Mac to the provider — VibeSpace runs no relay server and never touches your content.
API keys are encrypted with AES-256-GCM, kept apart from everyday config, live only in memory, and never enter logs.
Model costs bill straight to your own accounts with zero markup from VibeSpace — add, remove, or switch providers and models anytime.
What's inside
Turn work you've done once into a SKILL.md package — instructions, templates, and scripts, all reusable. Skills auto-match to new tasks, or you pick them by hand; every run follows the same standard.
First-class boards, columns, and cards that link to tasks and projects. Drag work across states and keep the whole picture in view.
One-time, daily, or weekly local schedules. Triggers fire while the app is open — each run gets its own traceable record, and risky actions still wait for your approval.
The life of a task
01 / Authorize & brief
Authorize a folder as the sandbox and state the goal in plain language. VibeSpace reads, writes, and runs commands only inside that boundary — anything beyond it is blocked.
Why does this module time out intermittently? Find the root cause first — don't change code yet.
I've traced the call paths and the latest performance logs.
02 / Visible execution
The timeline streams every file read, command, and code edit as it happens, with findings and diffs on record. Lose the network, crash, or close the app mid-run: tasks are durable and resume right where they stopped.
Every step stays attached to the task that started it.
03 / You decide
Overwriting files, risky commands, out-of-scope access — judgment calls pause for your approval. Every code change carries a reviewable diff and one-click restore, and history is never silently overwritten.
Write build output outside the authorized workspace (~/dist)?
This writes outside the workspace boundary. Nothing is written until you decide.
Built for real work
From the first question to the last scheduled run, context stays alive.
Reviewable changes, one-click restore.
Authorized workspace
This week's changes
Diff readyChangeSet
3 files · diff ready
Validation
Tests and build checks
Restore
Roll back any step
SKILL.md packages
PR #184 · src/session/context.ts
Skill matching
@@ -42,8 +42,13 @@ hydrateConversation
- return messages.slice(-20)
+ const history = await context.restore(threadId)
+ return compact(history, tokenBudget)
Skill resources
12Templates and scripts loaded with the skill
Run state
Ready0 failed checks
Evidence stays attached.
Input
Attachments
7 files
Open
02
Evidence
22 excerpts
Connected
03
Synthesis
3 findings
Draft
04
Output
1 document
Written
A rhythm you can inspect
Cadence
Upcoming runs
Repo weekly report
Dependency sweep
Progress digest
Weekly decisions
Reviewable changes, one-click restore.
Authorized workspace
This week's changes
Diff readyChangeSet
3 files · diff ready
Validation
Tests and build checks
Restore
Roll back any step
SKILL.md packages
PR #184 · src/session/context.ts
Skill matching
@@ -42,8 +42,13 @@ hydrateConversation
- return messages.slice(-20)
+ const history = await context.restore(threadId)
+ return compact(history, tokenBudget)
Skill resources
12Templates and scripts loaded with the skill
Run state
Ready0 failed checks
Evidence stays attached.
Input
Attachments
7 files
Open
02
Evidence
22 excerpts
Connected
03
Synthesis
3 findings
Draft
04
Output
1 document
Written
A rhythm you can inspect
Cadence
Upcoming runs
Repo weekly report
Dependency sweep
Progress digest
Weekly decisions
Security & privacy
Nothing leaves your machine unless you say so. That's not a setting — it's the architecture.
Tasks, run history, boards, and schedules all live in ~/.vibespace on your disk. No cloud, no uploads — ever.
API keys are encrypted with AES-256-GCM and kept apart from everyday config. They live only in memory and never enter logs or run events.
Least authority by design: out-of-workspace access, risky commands, and file overwrites all require an explicit approval card — and prompt injection inside documents can't grant anything.
Telemetry is off by default. From first launch, VibeSpace is a transparent tool that answers only to you.
Your workspace is just a directory — back it up, move it, audit it freely. Zero lock-in.
What's next
Shipped · v0.12
One-time, daily, and weekly local schedules — every trigger gets its own traceable record.
Next · v0.13
Built-in requirements, implementation, test, and review experts, plus delivery and review squads — multiple roles collaborating on one run.
Planned · v0.14
A stable Core host plus signed local module packs — capability grows, boundaries don't budge.
Goal · v1.0
Signed, notarized, and dependable.
Built to be trusted
“Crash, reopen, and the run picks up where it stopped. Wrong turn? One click restores the files. Out of bounds? It asks me first. Reliability isn't a spec line — it's the default behavior of every single run.”
Sarah Mitchell
Indie developer · Beta tester
100%
of runs are recoverable — state persists locally and resumes on reopen
0
duplicate side effects — idempotent tool calls never write twice
6
built-in tools, every call on record: list / search / read / edit / write / bash
Early users
“Monday reporting used to eat a whole morning. Now the scheduled run has already written the report before I open my laptop.”
Product manager · Beta user
“I turned our review checklist into a skill. It applies the same standard to every single run — no shortcuts.”
Tech lead · Beta user
“A run crashed overnight. Next morning it resumed from the checkpoint and didn't write a single file twice.”
Indie developer · Beta user
“Client files can never touch the cloud. It runs entirely on my Mac — that's the precondition for trusting it at all.”
Consultant · Beta user
“Every diff sits there waiting for my review before anything is committed. Once you've worked this way, there's no going back.”
Full-stack engineer · Beta user
“It knows exactly when to stop and ask me. That restraint is more reassuring than "fully autonomous".”
Operations lead · Beta user
FAQ
No. Tasks, run history, and config stay local in ~/.vibespace, with API keys sealed by AES-256-GCM. VibeSpace talks only to the model providers you configure — it has no upload channel of its own, and telemetry is off by default.
Eleven providers ship built-in: Claude, OpenAI, Gemini, DeepSeek, Kimi, Qwen, GLM, MiniMax, Doubao, and OpenRouter — plus any OpenAI-compatible custom endpoint. Paste an API key and you're connected. Configs are per-provider, switchable anytime, and keys stay encrypted locally.
Yes — with three layers of protection. First, changes stay inside the authorized workspace; anything beyond the boundary is blocked. Second, every change lands as a reviewable diff with a validation summary, committed only after your approval. Third, any change can be restored to the exact pre-edit files with one click. Overwriting your existing files always requires explicit approval.
A skill is a piece of Markdown with metadata (SKILL.md): when to use it, how to do it, and what the output looks like — optionally with templates and scripts. No code required. If you can write docs, you can write skills. Existing packages from the Claude and Codex ecosystems work as-is.
Browser assistants live in a tab and can't reach your file system. VibeSpace is a real desktop app: it reads and writes local files, runs commands, and manages boards and schedules — with every step traceable, interruptible, and recoverable.
Those are IDE/terminal tools built for developers. VibeSpace is a complete desktop product: usable without touching a terminal, with kanban, scheduled tasks, a skill center, and graphical approval flows built in. It's for everyone who wants to hand repetitive work to AI — coders and non-coders alike.
No. Schedules trigger only while the app is running — close it and they pause, with no catch-up for missed triggers. That's a deliberate design stance: everything is predictable and explainable, and nothing ever runs without your knowledge.
The desktop app is free. Model costs bill directly to your own API accounts — VibeSpace adds no markup and never touches your data.
Download VibeSpace, authorize a folder, and say what you need. In a few minutes you'll wonder how you ever worked without it.
Review this repo's changes from the last week and write a weekly report to docs/weekly.md
On it — reading the last 7 days of commits and changed files.
Report written to docs/weekly.md — 12 changes total, 2 breaking changes highlighted.
