🛡️ Vajra

The local-first AI agent workspace — your machine, your models, your data.

One ~9 MB Go binary that gives you a terminal UI, an offline web GUI and a REST API on top of local models through Ollama — with multi-agent swarm orchestration, your own files as context and a 4-layer anti-hallucination pipeline.

No API key. No telemetry. No cloud. Works fully offline.

Quick start · Screenshots · Agent swarm · Your files · Verification · Docs · FAQ

---

Vajra is a self-hosted AI agent workspace for people who don't want their prompts — or their files — leaving their computer. It runs a local model for you, wires five specialist agents and a review pipeline around it, and talks to you through a terminal, a browser or plain HTTP. Nothing is a service: the index is JSON, the sessions are JSON, the UI is embedded in the binary.

📸 See it

Every image below is the real application, captured on a real Ollama backend (llama3.1:8b).

_{The offline web GUI — vajra web, no CDN and no build step; everything ships inside the binary.}

_{Verified answers. Every reply can be graded by the 4-layer pipeline (here: ✓ verified · 97%).}

_{The agent bus — you can read every message the agents exchanged, live or re-opened later.}

The terminal UI ships in the same binary: Ctrl+T swaps to swarm mode, Ctrl+B opens the agent bus, Ctrl+V runs verification, Ctrl+E exports the session. Same agents, same pipeline, no browser.

✨ What's in the box

• 🧩 One static binary (~9 MB). No Python, no Node, no Docker, no pip install. Windows/macOS/Linux, amd64 + arm64.
• 🐝 A swarm, not a chatbot. A Chief of Staff plans your goal and 3–4 specialists work in parallel — then review each other and hand work off with @role: requests on a bus you can read.
• 🛡️ Four verification layers. Fast checks → embedding similarity → LLM-as-a-judge → human escalation. Grounded in the sources the model was actually given, and honest when there's nothing to check.
• 📎 Your files, verifiably local. Attach files (📎, drag-and-drop, @path, --file) or index a whole folder and ask questions with file:line citations — no vector database, no upload, no account.
• 🧮 Context that budgets itself. Long chats are trimmed to the model's window; the system prompt and your latest question always survive.
• 🖥️ Three interfaces, one core. TUI (vajra), web GUI (vajra web), REST + SSE (vajra serve) — or one-shot vajra run.
• ✍️ Export anything. Any session as Markdown or JSON, including the attachments used, the cited sources and the full swarm transcript.
• 🔌 Local-first ≠ local-only. Add an OpenAI/Anthropic/OpenRouter/Groq key only if you decide to — routing is automatic (claude-* → Anthropic, gpt-* → OpenAI, everything else → local Ollama).

🚀 Quick start

1. Get a local model (Ollama — one-time)

ollama pull llama3.1:8b

2. Install Vajra

Platform	Install
Windows	Download the VajraSetup-<version>-win-x64.exe installer from Releases
Linux / macOS	curl -fsSL https://raw.githubusercontent.com/Ritiksuman07/Vajra/main/scripts/install.sh | bash
Any (Go)	go install github.com/Ritiksuman07/Vajra@latest
From source	git clone https://github.com/Ritiksuman07/Vajra && cd Vajra && make build

3. Run

vajra          # interactive terminal UI
vajra web      # web GUI  →  http://127.0.0.1:4096
vajra doctor   # check your setup end-to-end

Every release ships with .sha256 checksums for each artifact — verify what you download with sha256sum -c.

🖥️ CLI in 60 seconds

vajra run "explain goroutines"                  # one-shot, streams to stdout
vajra run --agent coder "review this" < file.go # pick a persona, pipe input
vajra run --file notes.md "summarise this"      # attach local files
vajra run --rag ~/notes "summarise my retries"  # answer from an indexed folder
vajra run --swarm "design a CLI bookmark tool"  # multi-agent swarm
vajra index ~/notes                             # chunk + embed a folder
vajra rag ~/notes "what did I decide about retries?"
vajra verify "Paris is the capital of France."  # run the pipeline on any claim
vajra export [--format md|json] [--out path]    # export a session
vajra serve --port 8080                         # REST API + SSE

🐝 The agent swarm

Agent	Role	Suggested model
🧭 Chief of Staff	Plans, decomposes, delegates, synthesises	llama3.1:70b
💻 Coder	Writes and explains code	qwen2.5-coder:7b
🔬 Researcher	Analyses and summarises	llama3.1:8b
⚙️ Executor	Produces exact commands	phi3.5
🔍 Reviewer	Reviews for correctness & security	llama3.1:8b
🛡️ Assistant	Everyday general help	llama3.1:8b

The Chief plans the goal, the specialists make their round-1 parallel reply, the Reviewer posts a critique, and agents can ask each other for follow-ups with @role: request — parsed, dispatched and answered on the bus. Cost is bounded by design: one parallel round, one critique, at most three handoffs, one synthesis. The roster is never a lone agent — if a small model plans only one task, Vajra tops up from a default rotation so at least three distinct agents always run.

You can read the whole exchange while it happens (TUI Ctrl+B, web 🐝 panel) or after the fact — the transcript is stored with the session and exported to Markdown or JSON.

_{A real swarm run — goal in, synthesised plan out; the full agent transcript is stored with the session.}

📂 Your files, locally

Attach files — 📎 or drag-and-drop in the web UI (read in your browser, so the server never opens a path on your behalf), /attach or @path in the TUI, repeatable --file on the CLI. Binary and oversized files are skipped and reported — nothing disappears silently.

vajra run --file README.md --file internal/rag/rag.go "what does rag.go do?"

Ask a whole folder — index once, then search by meaning:

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 with a similarity score:

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

There is no vector database and no service: chunks are embedded by your local model, stored as JSON in ~/.vajra/index, and ranked with cosine similarity. Build output, dependency folders and dot-directories are skipped; file counts, chunk sizes and total index size are all bounded.

Long chats trim themselves. The conversation is budgeted against the model's context window (default 8192 tokens, VAJRA_MAX_CONTEXT to change). Older turns drop from the front; the system prompt and your latest question always survive. Both UIs show the level (ctx 2.1k/8.2k) and say so when something was trimmed.

✅ Anti-hallucination

Layer	What it does	Cost
1 · Fast checks	Structure, unbalanced code fences, PII, citation presence	~µs
2 · Semantic similarity	Embeds the answer + your sources, requires cosine ≥ 0.75	~50 ms
3 · LLM-as-a-judge	Rubric: correctness, completeness, faithfulness, coherence	~1 s
4 · Human escalation	Low-confidence answers are flagged, never hidden	instant

vajra run --verify "what is the capital of Australia?"
vajra verify "Paris is the capital of France."

Verification is grounded: your attached files and the passages retrieval cited are handed to layer 2 and the judge, so "verified" means the answer resembles its sources — not that it's a universal truth. When there was nothing to compare against, layer 2 reports that it was skipped rather than pretending to have checked.

Straight truth: these are heuristics, not proof. Vajra is upfront about it — the pipeline surfaces uncertainty visibly so you decide what to trust. Docs: how verification works.

🔐 Privacy in four lines

• Prompts go to localhost:11434. There is no telemetry, no account, no crash reporting — the binary makes no outbound calls unless you configure a provider.
• The web UI and API bind to 127.0.0.1 by default; --host is an explicit opt-in that prints a warning.
• Attachments in the browser are read by the browser and sent as text — the server never touches an arbitrary path on your behalf.
• Sessions, indexes and exports stay in ~/.vajra; vajra export writes them wherever you point it. Report vulnerabilities via SECURITY.md.

☁️ Optional cloud (as in, optional)

vajra connect openai     sk-...
vajra connect anthropic  sk-ant-...
vajra connect openrouter sk-or-...
vajra connect groq       gsk_...
vajra connect lmstudio           # local, OpenAI-compatible

Routing is automatic: claude-* → Anthropic, gpt-* → OpenAI, vendor/model → OpenRouter, everything else → local Ollama.

📊 How Vajra compares

	Vajra	Cloud chat apps	Other local agent frameworks
Works without an API key	✅	❌	✅
Runs fully offline	✅	❌	✅
Single binary, no runtime	✅	✅	❌
Terminal UI	✅	❌	⚠️ varies
Built-in offline web GUI	✅	✅*	❌
Multi-agent swarm mode	✅	❌	⚠️ varies
Readable agent-to-agent messages	✅	❌	❌
Verification pipeline	✅	❌	❌
Session export (md/json)	✅	⚠️ varies	⚠️ varies
Zero telemetry	✅	❌	⚠️ varies

* most "local" forks still rely on CDN assets or a build step.

🏛️ Architecture

Layered into focused packages under internal/ — the full tour is docs/ARCHITECTURE.md:

Package	Role
internal/config	Settings, VAJRA_* env precedence
internal/llm	Router (Ollama, OpenAI-compatible), token budgeting
internal/agent	Six roles, swarm orchestrator, message bus
internal/verify	The 4-layer anti-hallucination pipeline
internal/attach / internal/rag / internal/session	Files · local retrieval · sessions + export
internal/tui / internal/server	Terminal UI · REST + SSE + embedded web app

⚙️ Configuration

Config lives at ~/.vajra/config.json; sessions in ~/.vajra/sessions/, retrieval indexes in ~/.vajra/index/, exports in ~/.vajra/exports/.

vajra config --set model=llama3.1:8b
vajra config --set verify=true
vajra config --set ollama_host=http://localhost:11434

Env var	Effect
VAJRA_MODEL, VAJRA_AGENT, VAJRA_PORT, VAJRA_VERIFY	Defaults
VAJRA_MAX_CONTEXT	Context-window budget in tokens (default 8192)
VAJRA_CONFIG_DIR	Where config, sessions and indexes live
OLLAMA_HOST / VAJRA_OLLAMA_HOST	Ollama endpoint

📚 Documentation

Setup & troubleshooting	Install Ollama + Vajra, fix common issues
CLI reference	Every command, flag and keybinding
HTTP API	REST + SSE endpoints with examples
Agents	Roles, swarm mode, bus, routing
Architecture	How the pieces fit
FAQ	Privacy, specs, common questions
Roadmap	What shipped, what's next
Packaging	Installers & the release process

🤝 Community

Contributions are welcome — see CONTRIBUTING.md (make test runs go test ./... -race; CI runs vet, lint and the race detector on all three OSes and blocks on all of them). Good first tasks: good first issue, help wanted. Ask questions in Discussions.

Coming next: incremental re-indexing, tool calling behind an allow-list, and winget/Homebrew packages — see docs/TODO.md.

📄 License

MIT © Ritik Suman and contributors. See LICENSE. Cite Vajra via CITATION.cff.

⭐ Star history

If Vajra is useful to you, a ⭐ helps other people find local-first AI.

⬇ Download · 🐛 Report a bug · 💡 Request a feature · 💬 Discuss

_{local-first ai agent · offline ai · local llm · self-hosted · open source · ollama · golang · multi-agent · autonomous agents · anti-hallucination · privacy · privacy-first · terminal ui · web gui · rest api · zero telemetry · no api key · file rag · session export · model context budgeting}