Vajra v1.2.0

Vajra v1.2.0 — Swarm mode + your files as context 🐝

Vajra is a local-first AI agent workspace in a single Go binary. It runs
entirely on your computer through Ollama: a terminal UI,
a web GUI and a REST API, with multi-agent orchestration and a four-layer
verification pipeline. No telemetry. No account. No cloud.

✨ Highlights

• One binary, zero runtime. No Python, no Node, no Docker required.
• 🐝 Swarm mode — 3–4 specialists plan, work in parallel, review
  each other and hand work off with @role: requests. You can read every
  message they exchange, in the TUI (Ctrl+B) and in the web UI.
• 📎 Your files as context (new) — attach files by drag-and-drop, @path
  or --file, and answer questions about a whole folder with local retrieval
  that cites file:line.
• 🧮 Context budgeting (new) — long chats trim their oldest turns to fit the
  window, and both UIs show how full it is.
• 💾 Session export (new) — Markdown or JSON, including the attachments
  used, the cited sources and the full swarm transcript.
• Terminal UI with token streaming, session history, model/agent pickers.
• Web GUI (vajra web) — self-contained, offline, dark & light themes.
• 4-layer verification — fast checks, semantic similarity, LLM-as-a-judge, human escalation.
• Optional cloud — vajra connect openai (or Anthropic, OpenRouter, Groq, Together) if you choose.

📦 Install

Windows — download VajraSetup-1.2.0-win-x64.exe and run it.

Linux / macOS

curl -fsSL https://raw.githubusercontent.com/Ritiksuman07/Vajra/main/scripts/install.sh | bash

From source

git clone https://github.com/Ritiksuman07/Vajra
cd Vajra
make build

🚀 Quick start

ollama pull llama3.1:8b     # one-time model download
vajra                       # terminal UI
vajra web                   # web GUI in your browser
vajra run "explain goroutines"   # one-shot, streams to stdout
vajra doctor                # check your setup

📂 Your files, locally

Ask about your own files without them leaving the machine.

# attach files to one prompt
vajra run --file README.md --file internal/rag/rag.go "what does rag.go do?"

# or index a whole folder and ask questions against it
ollama pull nomic-embed-text      # a small embedding model
vajra index ~/notes
vajra rag ~/notes "what did I decide about retries?"
vajra run --rag ~/notes "summarise my retry decisions"

Answers cite what they used, as file:line:

[1] retries.md:1 (similarity 0.81)
# Retry policy
The retry policy uses exponential backoff with full jitter.

In the TUI: /attach <path>, @path/to/file, /rag <folder>. In the web UI:
📎 Attach (or drag-and-drop) and 📚 Folder, with live indexing progress.

• Nothing is uploaded. Files are read by your machine; retrieval is chunks
  embedded by your local model, stored as JSON in ~/.vajra/index, ranked by
  cosine similarity. There is no vector database and no service.
• The server is loopback-only. vajra serve binds 127.0.0.1 by default;
  --host opts out, with a warning.
• Long chats trim themselves. The context window is budgeted (default
  8192 tokens, VAJRA_MAX_CONTEXT); the oldest turns go first, your latest
  question always survives, and both UIs show ctx 2.1k/8.2k.
• Export anything. /export (TUI, Ctrl+E), ⬇ Export (web) or
  vajra export — Markdown or JSON, including attachments, sources and the
  full swarm transcript.

🐝 The agent swarm

Agent	Role
🧭 Chief of Staff	Plans, decomposes and delegates, synthesises the answer
💻 Coder	Writes and explains code
🔬 Researcher	Analyses and summarises
⚙️ Executor	Produces exact commands and file contents
🔍 Reviewer	Reviews for correctness and security

vajra run --swarm "add tests to a small Go project and review them"

The Chief plans, the specialists work in parallel, the reviewer critiques, and
agents can ask each other for follow-ups. Every message is shown with its
sender and recipient, so you can see the team think:

🐝 swarm plan · 3 agents
── round 1 — 3 agents working in parallel ──
  🔬 Researcher → 👥 everyone · reply
  💻 Coder → 👥 everyone · reply
  ✉ ⚙️ Executor → 💻 Coder   asked for help: also add tests for the parser
  💻 Coder → ⚙️ Executor · reply
── round 2 — the reviewer critiques the team's work ──
  🔍 Reviewer → 👥 everyone · critique
  🧭 Chief of Staff → 👤 you · synthesis

✅ Verification pipeline

Every answer can be checked locally before you trust it:

1. Fast checks — structure, PII, citation presence (~µs)
2. Semantic similarity — local embeddings vs. your sources (~50 ms)
3. LLM-as-a-judge — a rubric across correctness, completeness, faithfulness, coherence
4. Human escalation — low-confidence answers are flagged, not hidden

Layer 2 is grounded in what the model was actually given: your attached files
and the passages retrieval cited. If there was nothing to compare against, the
layer reports that it was skipped rather than pretending to have checked.

🔐 Privacy

Vajra sends nothing anywhere by default. Prompts go to localhost:11434.
The only outbound requests are ones you explicitly configure with
vajra connect.

📄 License

MIT © Ritik Suman and contributors.

Full release notes: see CHANGELOG.md

Vajra v1.1.0

Vajra v1.1.0 — Your machine. Your models. Your data. 🛡️

Vajra is a local-first AI agent workspace in a single Go binary. It runs
entirely on your computer through Ollama: a terminal UI,
a web GUI and a REST API, with multi-agent orchestration and a four-layer
verification pipeline. No telemetry. No account. No cloud.

✨ Highlights

• One binary, zero runtime. No Python, no Node, no Docker required.
• Terminal UI with token streaming, session history, model/agent pickers.
• Web GUI (vajra web) — self-contained, offline, dark & light themes.
• Multi-agent team mode — a Chief plans, specialists execute, the Chief synthesises.
• 4-layer verification — fast checks, semantic similarity, LLM-as-a-judge, human escalation.
• Optional cloud — vajra connect openai (or Anthropic, OpenRouter, Groq, Together) if you choose.

📦 Install

Windows — download VajraSetup-1.1.0-win-x64.exe and run it.

Linux / macOS

curl -fsSL https://raw.githubusercontent.com/Ritiksuman07/Vajra/main/scripts/install.sh | bash

From source

git clone https://github.com/Ritiksuman07/Vajra
cd Vajra
make build

🚀 Quick start

ollama pull llama3.1:8b     # one-time model download
vajra                       # terminal UI
vajra web                   # web GUI in your browser
vajra run "explain goroutines"   # one-shot, streams to stdout
vajra doctor                # check your setup

🧠 The agent team

Agent	Role
🧭 Chief of Staff	Plans, decomposes and delegates
💻 Coder	Writes and explains code
🔬 Researcher	Analyses and summarises
⚙️ Executor	Produces exact commands and file contents
🔍 Reviewer	Reviews for correctness and security

✅ Verification pipeline

Every answer can be checked locally before you trust it:

1. Fast checks — structure, PII, citation presence (~µs)
2. Semantic similarity — local embeddings vs. your sources (~50 ms)
3. LLM-as-a-judge — a rubric across correctness, completeness, faithfulness, coherence
4. Human escalation — low-confidence answers are flagged, not hidden

🔐 Privacy

Vajra sends nothing anywhere by default. Prompts go to localhost:11434.
The only outbound requests are ones you explicitly configure with
vajra connect.

📄 License

MIT © Ritik Suman and contributors.

Full release notes: see CHANGELOG.md