Skip to content
Tanvrit AI
Web App
Documentation

Developer documentation

Reference for every developer-facing surface in Tanvrit AI — install on macOS, Windows and Linux, wire up any MCP client, read the full 45-tool reference, and drive the same local runtime from the tanvrit CLI.

Run the 5-minute quickstart45-tool reference

Local search needs no signup — your code never leaves your machine.

Reference

Browse the docs

Every card links somewhere real. Topics whose long-form guide is still being written point at the nearest page that answers the question today.

Phase 2

Concepts

Local-first design, MCP protocol, indexer model, agent loop, conversation memory.

  • What Tanvrit AI is (and isn't)
  • Local-first vs cloud-first trade-offs
  • How MCP fits into your editor / agent
  • Indexer model: parsers, embeddings, SQLDelight schema
  • Agent loop: ToolDispatcher + ConversationSummarizer

Covered for now by the platform architecture overview.

Open

Installation

Get the desktop app, the VS Code extension, or the Wasm portal up.

  • macOS (Apple Silicon + Intel)
  • Windows MSI
  • Linux DEB
  • VS Code extension
  • Wasm portal at /app/ (limited features)
Open
Phase 2

Configuration

Settings file, bearer token rotation, telemetry, model selection.

  • Settings → MCP Bearer Token
  • Telemetry toggle (default OFF)
  • AI provider keys (Claude, OpenAI, Gemini, Ollama, …)
  • Indexer excludes and gitignore handling
  • Database location and SQLite inspection

The quickstart walks the settings you need on day one.

Open

Command line

The tanvrit CLI: the command tree, the --json envelope, exit codes and browser sign-in.

  • Two backends: attach to the running app, or run in-process
  • Full command reference, with what each backend actually serves
  • The --json envelope, worked examples and error codes
  • Exit codes 0 / 1 / 2 / 3 / 4 / 5 / 64
  • Browser sign-in (OAuth + PKCE) — implemented, blocked server-side
Open

MCP server

JSON-RPC over HTTP / SSE on localhost:19281, stdio mode for trusted parents.

  • Auth: Authorization: Bearer <token>
  • Endpoints: tools/list, tools/call, resources/list, prompts/list
  • Error model and retry semantics
  • Resource templates and contextual prompts
Open
Phase 2

Agent loop

How AgentLoop turns a user message into a sequence of tool calls.

  • Plan → call tools → summarise → answer
  • ConversationSummarizer for long sessions
  • AgentToolRegistry mapping
  • Stopping conditions and budget guards

The Agent Runtime pillar describes what ships today.

Open
Phase 2

Plugins

Extend Tanvrit AI with custom MCP tools and prompts.

  • Tool plugin contract
  • Prompt registration
  • Loading order and conflicts
  • Distribution (Phase 2)

Domain Packs are the shipped extension mechanism.

Open

API reference

Full schema for MCP tools, request / response shapes, error codes.

  • All 45 MCP tools (live list at runtime)
  • Request schemas (JSON Schema)
  • Response shapes
  • Error taxonomy
Open
Phase 2

Examples

End-to-end recipes: code review bot, refactor agent, dead-code finder.

  • Refactor agent with Claude Desktop
  • Dead-code finder via find_dead_code
  • Reviewer with analyze_impact + git_history
  • Custom HTTP client in 20 lines

The tool reference carries the call shapes each recipe uses.

Open

Transparency

Honest status

What ships in the product versus what is written down.

Sections marked Phase 2 ship in the product, but their long-form docs are still being written; each one links to the page that answers the question today.

The authoritative reference for tools is what your MCP client sees from tools/list at runtime.

Get started

Start with the quickstart

Install, index, wire up Claude Desktop, ask a question — in about five minutes. Your code never leaves your machine.

Run the 5-minute quickstartDownload