Common questions

Two components: ContextMint Engine (desktop app) and the ContextMint VS Code extension. Engine runs the Python server and Ollama; the extension is the chat and governance UI. See getting started and features (v1.0).

On the default local-first path, indexing, embeddings, retrieval, and inference run on your infrastructure via Ollama — source is not required to leave your device. If you opt in to BYOK cloud routing, only redacted assembled context is sent to a provider you choose; API keys stay in VS Code secrets and never pass through Engine. See Trust & Security.

Engine owns process lifecycle: server, Ollama, indexing operations, logs, and settings. ContextMint connects to localhost:8000, shows chat, Context Lens, quality map, and assisted actions — it does not spawn the server in packaged mode.

Local-first (default): open ContextMint Engine from the tray and use Start all. The extension probes health on loopback; if Engine is stopped, chat is gated until the backend is healthy again.

Shared on-prem server: confirm ContextMint.serverUrl points at your team API host, HTTPS/TLS is valid, and enterpriseMode + OIDC are configured if auth is enabled on the server. The extension does not install or start Engine in remote mode.

Yes — use the compose Context selector. Repo (default) answers only from your indexed codebase. Work is for planning, comms, and concepts using session packs and attachments without a repo file search. Hybrid combines repo evidence with reasoned recommendations. ContextMint is not a generic ChatGPT replacement; lanes keep repo answers trustworthy.

Chat is not blocked during background indexing — answers use whatever is indexed so far and improve as files finish. The readiness card may show an indexing banner; you can still send messages. If chat is fully blocked, check for server offline or index error state in Engine → Health.

Yes for local model inference. Engine’s first-run wizard detects Ollama and links to install if missing. If Ollama is offline, repo search may still work but AI generation runs in a degraded limited mode.

Context Lens is the pre-send preview: retrieved chunks, relevance scores, routing choice, and optional governance notes — so you approve evidence before the model runs. It is gate 1 in the human-in-the-loop workflow. See Context Lens & packs.

In chat compose, Ask scopes the model to answer your question, Plan asks for a structured plan before changes, and Agent allows broader multi-step assistance — still with Context Lens (gate 1) and patch preview (gate 2) before repo writes. Pair modes with Repo / Work / Hybrid lanes for where evidence comes from. See Chat, modes & lanes and Features (v1.0).

Assisted actions propose unified diffs; patch preview is gate 2 — you review and confirm before anything applies to your repo. Quality-map clone signals can block apply on high-severity findings when configured. No silent writes. See Patches & governance.

The quality map is stored governance findings from your repo at index time — clone pairs, near-duplicates, and structure signals. Chat can answer from this Findings Store (lookup) instead of rediscovering issues at question time. Inform-only by default; optional gates can block assisted apply. Run scans from the extension or Engine → Quality. See Knowledge & governance.

Context packs are per-repo knowledge bundles (manifest.yaml plus markdown). Activate with @pack:name in chat; the pack panel lists, builds, and auto-syncs when sources change. Packs feed Work and Hybrid lanes and appear in Context Lens before send. See Context Lens & packs.

ContextMint is a decision-control plane, not a speed-first autocomplete tool. It governs what evidence the model sees and what changes can apply, with audit-friendly telemetry. Human gates before inference and before apply matter more than raw generation throughput for regulated engineering workflows.

Yes — for enterprise pilots with platform engineering support. Deploy the API server stack (Docker Compose or Helm) on your infrastructure. Developers set ContextMint.serverUrl to that host and enable enterpriseMode for OIDC Bearer auth. The server must be able to read the workspace path being indexed — typically via VS Code Remote SSH, a mounted monorepo, or a dev VM where code and API run on the same host.

This is not the Engine desktop app (tray console is local-only). A laptop with a local-only clone cannot index through a remote server URL unless paths align. See Trust & Security → Deployment models.

ContextMint Engine is the per-machine operator console: tray app, first-run wizard, Ollama lifecycle, logs, and settings. It runs the API on loopback for individual developers — this is the v1.0 default.

A team API server is the same Python backend packaged for Docker/Kubernetes on customer infrastructure — shared Ollama, Chroma index, Redis rate limiting, and SSO. The extension becomes a thin client pointed at serverUrl. Ops owns deploy, mounts, and auth — not the Engine installer.

Yes — optional and off by default. Enable cloud routing in extension settings, store provider keys in VS Code secrets (OS keychain), and the extension streams directly to approved providers such as OpenAI, Anthropic, or Groq. Keys never pass through the Python engine. Assembled context is redacted before any cloud dispatch. The default path remains local Ollama on your infrastructure. See Trust & Security.

Yes. Paste, drag-drop, or pick PNG, JPEG, or WebP images in the compose bar and ask questions about what you see. ContextMint previews images in Context Lens and routes through a local Ollama vision model by default — no cloud API key required. Optional BYOK cloud vision is available when you explicitly select a vision model.

See Image evidence & ARGUS.

ARGUS is ContextMint’s structured visual audit on UI screenshots. From Context Lens, click Run visual audit to get severity-tagged findings (including WCAG references). ARGUS is report only — it never auto-fixes your repo. You can hand off to patch preview if you choose to act on a finding.

Sandbox audits (Playwright capture in preview worktree) are available when git sandbox is enabled.

It is a one-month evaluation, not a product subscription that expires. After onboarding, 10–30 developers on one team or repo use ContextMint in daily work for 30 calendar days. You give short async feedback at week 2 and week 4; we provide weekly check-ins and a written pilot summary for your security and platform stakeholders. Successful pilots can extend to 90 days. See Design Partners for the full breakdown.

v1.0 (pilot now — GA end of August 2026): local Engine + extension — Context Lens, Ask/Plan/Agent modes, Repo/Work/Hybrid lanes, Findings Store quality map, image evidence, ARGUS visual audit, context packs, assisted patch preview, glass-box routing, and symbol-aware citations. Full v1.0 list.

Later: v2.0 Engineering Intelligence Platform (Q1 2027), Track B enterprise knowledge and compliance (end 2027), and the long-term Verification Engine direction (PR gates, blast radius, drift) — full roadmap.

Design partner intake opens by start of September 2026 while we finish v1.0 pilot readiness. The form on Design Partners is visible for reference only — submissions are disabled until intake opens.

Available now: book a walkthrough via Book a Demo, try local-first install via Getting started, review program details on Early Access, or ask questions via Support or GitHub issues. When intake opens, design partners run a 30-day evaluation pilot (10–30 developers, one repo — extendable to 90 days).

Visit Support and open an issue on the release channel: https://github.com/contextmint/contextmint/issues. You can also email anis@contextmint.ai.