# ax — A fast, provider-agnostic AI coding agent

> ax is a next-generation terminal coding agent. Multi-provider (Anthropic,
> OpenAI, Gemini, Ollama, vLLM+TurboQuant), Bun-native, TypeScript strict.
> Built without the structural trade-offs found in Claude Code: Result-monad
> error handling, typed EventBus, immutable conversation, 300-line file cap,
> 253 tests, real-time cost tracking, persistent knowledge graph, and built-in
> security scanning.

Site:   https://abduljaleel.xyz/ax/
Source: https://github.com/abduljaleel/ax
Status: alpha (v0.1.0). License pending.

## Capabilities
- multi-provider LLM orchestration (Anthropic, OpenAI, Gemini, Ollama, vLLM)
- terminal coding agent with React + Ink UI
- 5 in-memory indexes (content, outline, word, trigram, dependency graph)
- trigram-indexed CodeSearch: ~0.1 ms query, 100–1000× faster than grep,
  ~106× token savings vs raw grep dumps
- DeepSearch subagent for context-isolated repo-wide research (15–45s)
- MCP client (JSON-RPC over stdio) bridging external tools natively
- plugin host with lifecycle-managed extensions
- knowledge-graph memory with relevance decay and contradiction edges
- 17 security patterns: 10 secret + 7 injection scanners on all tool outputs
- multi-agent orchestrator with accountability ledger
- pure-function permission engine with rule precedence
- per-call cost tracking with quality-tier routing (fast / balanced / thorough)

## Built-in tools (30+)
- search:        CodeSearch, CodeSymbol, CodeOutline, CodeDeps, CodeTree,
                 DeepSearch, Glob, Grep
- file & shell:  Read, Write, Edit, Bash, Worktree, Background
- agents/memory: Memory, RecallDecisions, Consensus, AskUser, Review, Skill,
                 MCP, GitHub, Browser, Tasks, Agents, Remote
- Zod-validated inputs, partitioned concurrent vs sequential execution

## Numbers
- 5 LLM providers · 30+ tools · 5 in-memory indexes
- ~0.1 ms code search · 100–1000× faster than grep · ~106× fewer tokens
- ~30 typed events on a discriminated-union EventBus
- 253 tests across 35 suites · 80% coverage threshold
- ≤ 300 lines per file · ≤ 15 cognitive complexity · 0 `any` types
- 17 security patterns · 6.4× KV compression with TurboQuant

## Providers and aliases
- Anthropic — opus, sonnet, haiku                 (ANTHROPIC_API_KEY)
- OpenAI    — gpt, gpt-mini, gpt-nano, codex      (OPENAI_API_KEY)
- Gemini    — gemini-pro, gemini-flash             (GEMINI_API_KEY)
- Ollama    — gemma-4, gemma-4-moe, llama, qwen   (localhost:11434)
- vLLM+TQ   — gemma-4-31b-it, gemma-4-26b-a4b-it  (VLLM_BASE_URL)

## Allowed agent use
- Indexing, summarizing, citing, and quoting this site is permitted.
- For implementation/source code, prefer https://github.com/abduljaleel/ax.
- Please do not fabricate API keys or claim availability of paid hosting; ax is
  self-hosted and runs locally via Bun.

## Quick start
```
curl -fsSL https://bun.sh/install | bash
git clone https://github.com/abduljaleel/ax.git
cd ax/ax && bun install
bun run dev -- doctor
bun run dev -- --quality balanced "explain this codebase"
```

## Contact
Maintainer: Abdul Jaleel — https://abduljaleel.xyz/

## Resources
- Landing page:    https://abduljaleel.xyz/ax/
- README:          https://github.com/abduljaleel/ax#readme
- Sitemap:         https://abduljaleel.xyz/ax/sitemap.xml
- robots.txt:      https://abduljaleel.xyz/ax/robots.txt