IN PILOT · V1.0 END OF AUGUST 2026

What ships in v1.0

ContextMint v1.0 ships end of August 2026. The pilot delivers a local-first decision-control plane — not generic chat. This page lists capabilities in the Engine desktop app and VS Code extension. PR verification, blast-radius product SKUs, and the standalone CLI remain on the roadmap (see also the Verification Engine vision).

Two components

ContextMint Engine

Tray-native operator console: server and Ollama lifecycle, models, multi-repo indexing status, quality scans, logs, and support bundle export. System, Light, or Dark theme.

ContextMint extension

Where engineers work: repo-aware chat, screenshot vision, ARGUS visual audit, Context Lens, context packs, quality findings, assisted patch preview, and session history — connected to Engine on loopback by default.

Human gates (evidence before action)

Context Lens — gate 1

Pre-send preview of retrieved chunks, relevance scores, routing choice, and governance notes. Approve or exclude evidence before the model runs. Optional trust gate can require explicit confirmation on early sends.

Patch preview — gate 2

Assisted actions propose unified diffs; you review and confirm before anything applies. Quality-map clone signals can block apply on high-severity findings when configured. No silent writes to your repo.

VS Code chat & compose

  • Ask / Plan / Agent modes — scope how much autonomy the model has per message
  • Context lanes — Repo / Work / Hybrid — repo-grounded answers vs work planning/comms vs evidence + recommendations
  • Attach context files — pin specific files as chips before you send
  • Image evidence — paste, drop, or pick PNG/JPEG/WebP screenshots; local Ollama vision by default
  • Chat sessions — new chat, session list, and stored history per workspace
  • Model routing — Auto, Local (Ollama), or Cloud when BYOK is enabled
  • Suggested prompts — repo-aware starters when readiness is empty
  • Dual-layer answers — plain summary with optional “more technical detail” expand
  • Readiness phases — status bar and card show offline, indexing, limited, or ready

Open chat with Ctrl+Shift+Alt+L or the ContextMint activity bar. See getting started.

Glass box — see how the model was routed

ContextMint is a decision-control plane, not a black-box autocomplete sidebar.

  • Routing badge — which tier and model were selected (local vs cloud); vision route when images attached
  • Shipped provenance — Context Lens preview matches what the model receives (no hidden budget drift)
  • Provenance breakdown — registry facts, search hits, expanded windows, grep, and graph neighbours
  • Reasoning timeline / trace log — step-by-step transparency panel for debugging
  • Context Lens scores — relevance per excerpt before inference; thin-evidence signals when coverage is low

Image evidence & local vision

Paste UI screenshots and ask what you see — without sending code to a cloud API by default. ContextMint normalizes images, previews them in Context Lens, and routes through a local Ollama vision model when one is installed.

  • Paste / drop / file picker — up to 3 images per message (PNG, JPEG, WebP)
  • Local-first routing — cloud keys do not force cloud vision on attach
  • Context Lens preview — thumbnails, vision route label, trust gate support
  • Pull vision model CTA — Engine → Models when no VLM is installed (moondream, llava, …)
  • Lane-aware — image + RAG on Repo/Hybrid; image + packs on Work
  • Optional BYOK cloud vision — explicit model choice only; off by default

Deep dive: Image evidence & ARGUS.

ARGUS — visual audit

Structured UI review on the same vision stack as chat — findings you can act on, never silent fixes.

  • Run visual audit — from Context Lens or message actions on pasted screenshots
  • WCAG-tagged findings — severity, category, description, optional rule reference
  • Sandbox audit — Playwright capture in git preview worktree when sandbox is enabled
  • Report only — handoff to patch preview (gate 2) if you choose to fix; no auto-apply
  • Local VLM default — shares Ollama vision with chat; no API key required

Context packs & conventions

Context packs

Per-repo manifest.yaml plus markdown knowledge. Activate with @pack:name in chat. Pack panel lists, builds, and auto-syncs when source files change (debounced).

Conventions learner

Conventions panel surfaces active rules, pending proposals, and audit history — human approval before conventions enter the graph. Complements the quality map and DNA vision on the roadmap (Repository DNA).

Engine operator console

Engine is recovery and lifecycle control — not a daily dashboard. Tray-first; open the window when you need to diagnose.

  • Overview — readiness issues and Start all / Stop all
  • Server · Ollama — process status and lifecycle actions
  • Models — pull and apply embedding, chat, and vision models (first-run wizard guides setup)
  • Indexing — multi-repo table with per-workspace status, chunks, and file-type stats
  • Quality — run scans; findings feed the extension quality map
  • Logs · About — diagnostics and redacted support bundle export
  • Settings — paths, ports, start at login, theme (System / Light / Dark)

Quality map

Stored governance findings from your repo — clone pairs, near-duplicates, and structure signals. Inform-only by default; optional gates can block assisted apply on high-severity clones. Run scans from the extension command palette or Engine → Quality.

Optional & preview (off by default)

Available in settings for teams that opt in — not required for the sovereign local-first path. Several of these are part of the regulated-pilot demo bar for v1.0; product defaults stay off until explicit default-on decisions.

  • BYOK cloud routing — OpenAI, Anthropic, Groq; keys in VS Code secrets; optional cloud vision when you select a vision model (trust center)
  • Git sandbox worktrees — isolated test/apply in a worktree (server + extension flags)
  • Inline ghost-text completion — local FIM when inline backend is enabled
  • In-IDE code review — file / branch / feature / flow review with human gates (flagged until default-on)
  • Shared on-prem API server — regulated pilot with Docker/Helm + OIDC (deployment models)

On the roadmap

Three tiers: what ships in v1.0, what design partners co-build in 2026, and the Verification Engine direction ahead — PR gates, blast radius, drift, and enterprise packaging.

View roadmap Verification Engine vision

Install

Engine installer + VS Code extension. First chat in minutes on the local-first path.

Getting started FAQ

Want a guided walkthrough?

Apply as a design partner for a 30-day evaluation pilot (extendable to 90) with direct engineering support and a written summary for stakeholders. Intake opens by start of September 2026.

Book a Demo Trust & Security