The Private AI Engineering Platform
Tanvrit AI is the private AI engineering platform — and with v1.5.0, an agent platform: named Engineering Agents, an in-app Marketplace, and Domain Packs that operate real business systems, all on six pillars built on one local core. Your code never leaves your machine; teams share curated knowledge, never code. Below: what ships today (desktop v1.5.26, VS Code extension v1.5.1 — both live) and what is still rolling out on the cloud side.
How it fits together
Everything inside the solid boxes runs on your machine. The dashed box is optional, opt-in, and never sees your code.
Code and the index stay on the left of the dashed line. The only things that ever cross it — and only if you opt in — are gateway-routed LLM calls, curated team knowledge, marketplace installs, and scheduled agent jobs. Direct-to-provider calls with your own keys remain the default.
The six pillars
AI Router
ShippedEvery agent request goes through one routing layer instead of a hard-coded provider, so a rate limit or outage never kills a run.
- ✓7 providers behind one interface — Anthropic Claude, OpenAI, Gemini, DeepSeek, Groq, Mistral, Ollama, and local engines
- ✓Automatic fallback — on rate-limit or network errors the run continues on the next provider in your priority list
- ✓Circuit breakers — a failing provider is taken out of rotation and re-probed, instead of being hammered
- ✓Cost-aware model registry — pricing and capabilities for ~25 models drive routing decisions and the cost dashboard
- ✓Org routing policy enforced locally — model and provider allowlists plus local-only mode, applied once your org's policy syncs after sign-in
- ✓AI gateway live — org policy distribution and team usage metering run through the Tanvrit gateway today
- →Hosted inference — Tanvrit-managed vendor keys on the gateway, so teams can route to cloud models without bringing their own. Bring-your-own-key routing works today
Agent Runtime
ShippedThe agent chat grew into a full runtime in 1.4.0 — and into an agent platform in 1.5.0: named agents, a curated marketplace, and domain packs on the same staged, verified, attributed runtime.
- ✓Four named Engineering Agents (1.5.0) — Code Reviewer, Test Writer, Migration, and Docs, each 1.0.0, documented, with a per-agent tool scope enforced by the runtime (the Docs agent can physically only write Markdown)
- ✓Agent Marketplace (1.5.0) — browse and install curated agents and Domain Packs in-app; payloads are validated, built-in ids can't be shadowed, and every install is provenance-stamped
- ✓Domain Packs + Commerce pack v1 (1.5.0) — 12 governed commerce_* tools over the /api/ai/v1 gateway plus the Store Operations Agent; opt-in per pack, every mutating call approval-gated
- ✓Staged edits — the agent proposes diffs; you apply or reject every hunk before anything touches disk
- ✓Persistent run history — tokens, cost, edit count, and source (UI / API / VS Code / Compute) per run, now with configurable retention (1.5.0)
- ✓Headless CLI runs (tanvrit-ai-agent) — plan-only by default; --auto-approve stages, verifies, and rolls back on failure; scheduled agents on Tanvrit Compute, cron included
- ✓Bearer-hardened Local API (1.5.0) — agent endpoints now require the same bearer token as the rest of the surface
- →Open marketplace publishing — third-party agents with review, signing, and revenue share; curated first-party listings only today
- →Education and DevOps Domain Packs — the commerce pack is the proof; more business systems follow
Repository Intelligence
ShippedThe local index is the platform's foundation — symbols, graphs, and search computed from your real code, on your machine.
- ✓19-language ANTLR-based indexer with incremental, hash-based re-indexing
- ✓Dependency and call graphs computed from real imports and references — not filename guessing
- ✓Hybrid search (since 1.4.0) — FTS5 full-text fused with real semantic embeddings via reciprocal-rank fusion (local engine or Ollama nomic-embed-text)
- ✓~45 built-in MCP tools on localhost:19281, plus per-pack tools that register when you enable a Domain Pack — chunked retrieval saves 90–95% of tokens per AI call
Team Knowledge
Shipped — opt-inTeams accumulate decisions, rationales, and annotations on top of code. The platform shares that knowledge — and only that knowledge.
- ✓Local annotation layer — notes attached to symbols and files, stored in your local index (shipped in 1.1.0)
- ✓Opt-in sync of curated decisions, rationales, and annotations across your team — default OFF, and every payload runs through the privacy redactor
- ✓Team Knowledge screen, plus /v1/knowledge/search on the Local API (powers the VS Code knowledge view)
- ✓Hosted team-knowledge backend — live; sharing activates the moment your team opts in
- ✓Hard boundary by design: the raw code index never leaves your machine — what syncs is what you explicitly curate, never code
Enterprise
Server-side liveControls for organizations that need governance without giving up the local-first core.
- ✓BYOK provider keys, stored encrypted at rest in your platform keychain
- ✓The app runs fully local by default — no account or cloud required
- ✓Org model and provider allowlists, enforced in the client; enforced local-only mode locks routing to local providers
- ✓Team usage dashboard, metered through the live AI gateway
- ✓Seats and entitlements — live on the Tanvrit server
- ✓Marketplace curation flow (submit → review → approve) — live; only approved listings are installable
- ✓Self-host bundle — in-repo preview that runs the full stack on your hardware in evaluation mode
- →Inbound enterprise SSO
- →License enforcement for the self-host bundle — evaluation mode today
VS Code Copilot Experience
LiveExtension v1.5.1 is live on the VS Code Marketplace and pairs fully with desktop v1.5.26. Note: extensions 1.3.x and older can no longer run agents against desktop 1.5.26 (agent endpoints now require the bearer token); upgrade to 1.5.1.
- ✓Streaming agent runs in a dedicated panel, with live tool-call output
- ✓Pending Edits review — apply or reject each staged hunk from the editor; the extension never writes files itself, the desktop app applies what you approve through its pending-edits API
- ✓Editor context sync — your active file and selection flow into agent runs
- ✓Agent run history, browsable from the sidebar
- ✓Knowledge view, backed by /v1/knowledge/search on desktop 1.4.0+ — fills in from the live team-knowledge backend when your team opts in
Privacy is the architecture, not a setting
The index is a SQLite database on your disk. Parsing, dependency and call graphs, full-text search, and all MCP tool execution run 100% locally, behind a bearer-token-gated server on localhost. When teams collaborate, only curated knowledge — the decisions, rationales, and annotations you explicitly choose to share — syncs through Tanvrit cloud; the raw code index never leaves your machine. Provider keys are bring-your-own and encrypted at rest, and enforced local-only mode for orgs cuts outbound traffic entirely.
Your code never leaves your machine; teams share curated knowledge, never code.
Honest: what you can install today
Desktop v1.5.26 and the VS Code extension v1.5.1 are both live and pair fully — everything marked “Shipped” above is installable today, including the four Engineering Agents, the Agent Marketplace, and the Commerce pack. Server-side, the AI gateway, marketplace curation, seats/entitlements, and the team-knowledge backend are live. The self-host bundle is a preview: it ships in-repo and runs the full stack in evaluation mode — license enforcement is on the roadmap. The one capability still rolling out is hosted inference (Tanvrit-managed vendor keys); bring-your-own-key routing works today.