Mitii AI Agent

Local-first AI coding agent for VS Code. Bring your own model (Ollama, vLLM, any OpenAI-compatible endpoint), keep your code on your machine, and get an agent that actually knows your repo before it starts editing files.

Docs: docs.mitii.dev · Discord: discord.gg/sa8rubf6HH

Built by codewithshinde.

The problem

Most coding agents fall into one of two camps:

1. Cloud-only — fast models, but your source code leaves the machine and you have limited control over what the agent can touch.
2. Local but blind — the model runs nearby, but context is shallow: a few @-mentioned files, maybe a grep, then a patch that misses half the codebase.

Both approaches break down on real work. Refactors span dozens of files. Audits need dependency graphs, not just string search. Long sessions blow past context windows. And when an agent writes to disk or runs shell commands, you want guardrails — not a binary "trust everything" switch.

Mitii sits in the middle: local execution, deep repo indexing, explicit Plan/Act separation, and approval policies you can tune from "ask me about everything" to "just run it, but block rm -rf".

How Mitii solves it

Features

Context and indexing

• Workspace scanner respects .gitignore and .thunderignore; auto-indexes on folder open
• FTS5 full-text search with ripgrep fallback for unindexed paths
• Symbol extraction for TypeScript, JavaScript, Python, Java, Go
• PageRank repo map — surfaces the files that matter structurally, not just lexically
• Vector search — local MiniLM embeddings via @xenova/transformers, hash fallback if unavailable; SQLite or LanceDB backend
• Hybrid retriever + reranker — merges search, vectors, mentioned files, git state, and LSP diagnostics into one context pack

Agent workflow

• Plan / Act / Review modes — plan before you touch code; act with tools; review without rewriting
• Tool loop — read_file, search, write_file, apply_patch, run_command, git_diff, diagnostics, and more
• Research subagents — parallel read-only workers for exploration (spawn_research_agent)
• Task decomposition — multi-step plans with lifecycle tracking and step completion markers
• Post-edit verification — configurable lint/test commands after Act-mode runs
• Skills catalog — drop SKILL.md files in .thunder/skills/ and invoke them via use_skill

Safety and control

• Approval modes — review_all, ask_edits, ask_deletes, ask_commands, auto
• Autonomy presets — safe, guided, builder, pilot, enterprise (network disabled)
• Untrusted workspace blocking — writes and shell disabled unless you explicitly opt in
• Auto-checkpoints before approved file writes
• Optional diff preview in VS Code before patches land
• Patch validation — syntax guards before apply; refuses shell commands masquerading as source code

Memory and persistence

• Long-term memory — memory_search / memory_write with FTS5 + optional vector hybrid search
• Post-task memory extraction — observations captured after completed work
• Session history — agent_sessions and agent_turns stored in SQLite; resume from the History tab
• Plan persistence — plans saved to task_plans and .thunder/tasks/

MCP and integrations

Preloaded keyless MCP servers (disable with thunder.mcp.preloadBuiltin: false):

Add your own via thunder.mcp.servers, .thunder/mcp.json, or .mcp.json. MCP tools appear as mcp__server__tool and still pass through Mitii's approval policy.

UI

React sidebar webview with chat, history, settings, approval cards, subagent activity panel, plan panel, checkpoint browser, indexing status, context budget warnings, and token meter.

Quick start

Requirements: VS Code 1.85+, Node.js 20+

git clone https://github.com/codewithshinde/thunder-ai-agent.git
cd thunder-ai-agent
npm install
npm run compile

Press F5 in VS Code to launch the Extension Development Host. Open a folder, wait for indexing to finish (status bar in the Mitii sidebar), then chat.

Connect a model

1. Open Settings in the Mitii sidebar (or VS Code settings under Mitii AI Agent)
2. Set thunder.provider.type to openai-compatible
3. Point thunder.provider.baseUrl at your endpoint (default: http://localhost:11434/v1 for Ollama)
4. Set thunder.provider.model (default: qwen3-coder:30b)

Use the Echo provider for UI testing without an LLM. API keys go through VS Code SecretStorage via the settings UI.

Commands

Configuration highlights

{
  "thunder.provider.type": "openai-compatible",
  "thunder.provider.baseUrl": "http://localhost:11434/v1",
  "thunder.provider.model": "qwen3-coder:30b",
  "thunder.safety.autonomyPreset": "guided",
  "thunder.safety.approvalMode": "review_all",
  "thunder.indexing.autoIndexOnOpen": true,
  "thunder.indexing.vectorsEnabled": true,
  "thunder.agent.verifyCommands": ["npm run lint", "npm test"],
  "thunder.telemetry.sessionLogging": true
}

See package.json → contributes.configuration for the full schema, or the Settings tab in the sidebar.

Project rules

Mitii picks up methodology files automatically:

• AGENTS.md, CLAUDE.md, WARP.md, .cursorrules
• .thunder/rules, .thunder/agents, .thunder/checks, .thunder/prompts
• .clinerules, .continue/rules, .cursor/rules

Commit these to your repo so every session starts with the same conventions.

Architecture

VS Code Extension
  └─ ThunderController
       ├─ SQLite Index (FTS5, symbols, vectors, sessions, memory)
       ├─ HybridRetriever → ContextBudgeter
       ├─ ChatOrchestrator
       │    ├─ AgentLoop (tool round-trip)
       │    └─ PlanExecutor (step engine)
       ├─ ToolRuntime → ToolPolicyEngine → ApprovalQueue
       ├─ PatchApplyService / CheckpointService
       ├─ MemoryService + PostTaskMemoryWorker
       └─ McpManager

Data lives in .thunder/ inside your workspace (thunder.sqlite, logs, checkpoints, tasks). Nothing is sent to a Mitii server — there isn't one.

Troubleshooting

better-sqlite3 fails to load

VS Code and Cursor ship their own Electron runtime, not your system Node:

npm run rebuild:native          # VS Code (auto-detects Electron)
THUNDER_EDITOR=cursor npm run rebuild:native   # Cursor
npm run rebuild:node            # for local vitest runs

Provider errors — check base URL and model name. Echo provider works without a running LLM.

Indexing stuck or empty — verify the workspace is writable, check .gitignore / .thunderignore, run Mitii: Index Workspace.

Context feels thin — confirm indexing finished, enable vectors (thunder.indexing.vectorsEnabled), and check the context budget warning banner in chat.

Related repositories

Scaffold copies live in mitii-docs/ and mitii-website/ at the repo root for convenience. Each is intended to be its own git repository — see the README in each folder for git init and deploy steps.

Development

See CONTRIBUTING.md for setup, project layout, testing, and pull request guidelines.

npm run watch     # extension + webview hot rebuild
npm run test      # unit tests
npm run lint      # typecheck
npm run package   # build .vsix

Author

codewithshinde
GitHub: @codewithshinde
Email: codewithshinde@gmail.com

Questions, bug reports, and PRs welcome on GitHub Issues.

License

Mitii AI Agent is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0-or-later).

If you run a modified version as a network service, AGPL requires you to make the corresponding source available to users of that service. For commercial licensing outside AGPL terms, contact codewithshinde@gmail.com.