Meta AI chat
Center-stage chat on Muse Spark 1.3 (default) or Muse Wonder 1.0, with workspace context, research-brief context, history, and Stop/cancel.
Case study · KryomLabs
KryomCode
A Meta AI-powered desktop IDE that bridges scientific research and software engineering — where every AI-assisted change is proposed, reviewed, approved, and verified, with the evidence kept alongside the code.
- Version
- 0.1.0 (MIT)
- Released
- 2026-11-03
- Platform
- Windows x64 installer
- Role
- Solo developer
- Team
- KryomLabs #1011
- Track
- Indie Developer (Solo)
An independent KryomLabs project integrating Meta AI models through their API; this does not imply official endorsement by, or affiliation with, Meta. Facts on this page are verified against the public repository README.
Overview
Review-first companion for evidence-backed development
KryomCode is an AI-assisted coding and scientific research desktop workspace. It keeps the whole development chain in one Meta-first environment: research with cited sources, source-grounded plans, diff review, explicit approval, then verified execution.
It does not claim to replace every IDE or coding agent. Instead, it is the place where AI assistance is always proposed, reviewed, approved, and verified — with the evidence kept alongside the code.
Problem & motivation
Research dies in browser tabs
Papers get read, then implementation starts from a blank file with no trace back to the evidence. Meanwhile AI coding tools either act without review or bury their reasoning where it cannot be checked.
The result is code nobody can audit: no sources, no plan record, no proof that verification actually ran. KryomCode exists to close that gap.
Vision
Accessible, transparent, verifiable
KryomCode is built toward a workspace where a single builder can move from a research question to reviewed, verified implementation without losing the chain of evidence — and where every AI suggestion arrives with its reasoning visible and its execution gated behind human approval.
The long-term direction is evidence-driven AI-assisted engineering and scientific research. Intended application areas include robotics, quantum computing, scientific computing, machine learning, and aerospace — directions for the product, not claims that every specialized workflow is already implemented.
Design philosophy
- Proposed, never presumed — the agent suggests, the human decides.
- Reviewed with intent — every edit carries its rationale and its diff.
- Verified with confidence — reports pin to observed runs, not model prose.
- Honest by default — no silent fallbacks, no invented model properties, no faked checks.
How it works
Research-to-engineering workflow
One fixed pipeline from question to verified code. Each stage hands evidence — not just output — to the next.
- 01ResearchInvestigate through real search providers with cited briefs.
- 02UnderstandIndex the codebase — symbols, impact, architecture.
- 03PlanAgent proposes a bounded plan grounded in sources.
- 04ReviewEach edit shows rationale plus diff for approval.
- 05ExecuteApply only selected edits, with rollback on failure.
- 06VerifyGuarded runs produce reports pinned to observed runs.
Under the hood
System architecture
Three layers, one trust direction — the renderer never touches disk, network, or secrets. Electron 45 · React 19 · TypeScript 5.7 · Vite 6.
Renderer
React UI shell · sandboxed
- App shell + workbench views
- Monaco Editor (local bundle)
- xterm.js transcript renderer
- Typed kryom() bridge access
Preload bridge
window.kryom · no Node APIs
- ~200 typed capability methods
- Sender + payload validation
- Human-readable messages only
Main process
Node + Electron · services
- 15 IPC domain modules
- workspace · terminal · ai · agent · git · research…
- Disk, Git, Meta AI endpoint access
Trust notes
- One trust direction: the renderer never touches disk, network, or secrets.
- Every path resolves through workspace containment — traversal rejected, symlinks refused.
- Per-project state lives in .kryom/ (git-ignored, never committed).
- Keys and secrets stay in the main process; only presence booleans cross IPC.
Drawn from the repository's documented architecture. Full details live in the project docs — see the links section below.
What it does
Core capabilities
Implemented and evidence-backed — each capability below is documented in the public repository.
Approval-gated agent
Task to validated plan to per-edit approval, with atomic apply, rollback, checkpoints, per-item undo, and freshness protection.
Research workspace
Visible run steps, persisted sessions, cited markdown briefs, source lists, .docx export, and Implement-from-Research handoff.
Evidence and citations
Citations must resolve to retrieved sources; claim checks are lexical and labeled as such; model prose is marked AI-generated.
Diff review and rollback
Why-rationale plus diff per edit, change-set ordering, same-file conflict blocking, and digest-checked undo of pre-apply bytes.
Monaco editor
Multi-tab editing with dirty tracking, session restore, conflict dialogs, auto save, and locally bundled workers (no CDN).
Integrated terminal
One explicit run per command with timeout, output caps, secret scrubbing, advisory warnings, and cancel/status controls.
Git source control
Status, diffs, stage/unstage, confirmed discard, commit, push, fast-forward-only pull, and restricted clone sources.
Verification reports
Reports derive from observed run facts — exit codes and run ids — never from model prose. Stale steps are flagged, never faked.
Built with
Technology stack
- Electron 45
- Desktop shell, lifecycle, main process
- React 19 + TypeScript 5.7 (strict)
- Renderer UI
- Vite 6
- Renderer build and dev server
- Monaco Editor
- Locally bundled editing surface
- xterm.js
- Buffered transcript renderer (not an interactive PTY)
- node:test + Vitest
- Backend suite and renderer suite
- electron-builder 26
- Packaging and installers
- Fixed Meta AI endpoint
- api.meta.ai chat completions · temperature 0.2 · 90 s timeout
Research integrations
- arXivKeyless
- OpenAlexKeyless
- CrossrefKeyless
- Semantic ScholarKeyless
- DuckDuckGoKeyless
- TavilyOptional · API key
Claim verification uses lexical matching and is labeled as such — the project does not claim semantic search for these integrations.
Human approval & safety model
Plans propose, humans approve
Every file edit needs explicit selection; apply, run, skip, and undo each confirm first.
Destructive commands need native approval
Main-process approval for destructive terminal commands that a compromised renderer cannot skip.
Checkpoints before every apply
Pre-apply bytes persist with digest-checked undo; failures roll back with honest counts.
Sandboxed renderer
Context isolation, no Node integration, full sandbox; navigation restricted; CSP allows no remote hosts.
Secrets stay sealed
Environment allowlist plus OS-encrypted vault; terminal and Git children get scrubbed environments.
Bounded recovery
Failed checks enter a bounded repair cycle (2 attempts, 10-minute budget) with per-attempt approval.
How it shipped
Development journey
Milestones with documented dates. Anything undated stays unspecified.
- 20262026Hackathon entryEntered in the Meta Global AI Developer Hackathon 2026, Indie Developer (Individual/Solo) track — Team KryomLabs #1011, problem statement “Push the Frontier”. Solo developer.
- 2026-11-022026-11-02Demo recordedOfficial 2:16 hackathon demo recorded following the published demo script.
- 2026-11-022026-11-02Meta AI live verificationOne exact-marker request per model, both passed — Muse Spark 1.3 and Muse Wonder 1.0.
- 2026-11-032026-11-03CI green across three OSesAll six jobs passed: test and build plus pack-and-verify on macOS, Ubuntu, and Windows.
- 2026-11-032026-11-03Owner GUI smoke passAll seven manual checks passed on the installed Windows app: launch, Explorer, chat, agent, terminal, Git, research-to-agent.
- 2026-11-032026-11-03Public release v0.1.0MIT-licensed release with the install-tested Windows x64 installer (KryomCode.Setup.0.1.0.exe, ~143 MB).
Real captures
Screenshots
Authentic captures from the running app (dark theme), served from the project repository with their documented captions.
01 · Main workbench
Welcome view with Chat/Research/Agent modes, Explorer, prompt box, and per-request model picker.
02 · Meta AI chat
Live request with elapsed timer and Stop, workspace context chips, and filterable thread history.
03 · Agent planning
Staged execution plan with per-stage files and Verify commands, explicit approval gate, and live run timeline.
04 · Research workspace
Persisted sessions, validated multi-query planning progress, and the Overview-to-Report tab flow.
05 · Research evidence and citations
Confidence-graded claims with numbered citations and one-click handoff briefs for chat and the agent.
06 · Integrated terminal
One-shot reviewed execution with Run/Cancel/Status/Clear/Copy, above the Platform workflows tab.
07 · Git source control
Staged/unstaged/untracked, commit, Push/Pull, plus the Tasks panel with plan verification states.
08 · Meta model configuration
Settings picker with both Muse models; API keys are managed under Providers and never shown.
Source: KryomCode repository docs gallery. Additional supplementary shots stay in the repository outside this gallery.
In motion
Demo video
The official 2:16 hackathon demo, recorded 2026-11-02 following the published demo script.
H.264 1080p · AAC stereo · 2:16
Watch on YouTubeGet it & follow it
GitHub and releases
Honest scope
Known limitations
- Windows is the validated platform; macOS and Linux packaging is configured but runtime there is not fully owner-tested.
- Evidence retrieval uses lexical matching — claim checks are labeled lexical, not semantic search.
- The installer is unsigned, so Windows SmartScreen may warn on first run.
- The terminal renders buffered transcripts; it is not an interactive PTY.
- Plans are bounded: at most 20 file edits and 10 suggested commands per plan.
- Files are capped at 1 MB in the editor; prompts are capped at 12,000 characters.
What is next
Future roadmap
Three-OS CI matrixDone
First full matrix runs recorded — completed 2026-11-03.
macOS and Linux hardware validation
Build, launch, and smoke-test installers on real machines. Blocked: test machines.
Signed installers
Windows certificate plus Apple notarization. Blocked: certificates.
Hosted update feed
Automatic update channel. Blocked: hosting.
Streaming output and more parsers
Ideas only — no commitments or schedule.
Roadmap order follows the repository; nothing beyond the completed item is scheduled or promised.