exara-agent 0.1.0

Local-first autonomous AI agent for programming and system automation. Runs against local models (Ollama, vLLM, LM Studio) or any OpenAI-compatible API (OpenRouter, OpenAI, Anthropic-compatible) with native tool calling, MCP support, plan mode, and skills.

exara-agent

Local-first autonomous AI agent for programming and system automation. Runs against local models (Ollama, vLLM, LM Studio, llama.cpp) or any OpenAI-compatible API (OpenRouter, OpenAI, DeepSeek, Anthropic, Groq) using native tool calling. Ships with a Rich CLI, a Next.js web UI, and MCP (Model Context Protocol) support so you can plug in GitHub, filesystem, Slack, Linear, Postgres, and the rest of the community catalog.

pip install exara-agent
exara init      # one-time interactive setup
exara chat      # talk to it from any folder

---

What it does

• ReAct loop with native tool calling — no fragile JSON parsing, the model emits real function calls.
• 30+ built-in tools — file ops, multi-edit, search, shell, background processes, Python execution, git, package install, web fetch, vision, audio, todos, plan mode, subagent delegation.
• MCP client — ~/.ai-agent/mcp.json works the same as Claude Desktop / Cursor: drop in any server from the catalog and the tools appear in the agent. Multi-source merge: bundled defaults → user-global → per-workspace.
• Skills system — markdown knowledge packs with auto-activation by stack (Python, Next.js, FastAPI, …). 10 ship out of the box; add your own at ~/.ai-agent/skills/ (global) or ./skills/ (per-project).
• Multi-provider profiles — exara profile use openrouter to swap. Keys live in ~/.ai-agent/profiles.json, never in the repo.
• Three-tier permission model — safe / normal / full, with a hard denylist for destructive commands (rm -rf /, mkfs.*, format c:, …) enforced at every level.
• Streaming responses + per-tool diff previews — Claude Code-style UX.
• Auto-context compaction — keeps long sessions inside the model's window without you noticing.
• Hooks — shell commands fired on session_start, pre_tool_use, post_tool_use, etc. Same model as Claude Code.
• Docker sandbox (optional) — route run_python and execute_terminal through a disposable container instead of your host shell.
• Secret redaction — OPENAI_API_KEY=…, GitHub PATs, AWS keys, Bearer tokens, and DB URLs are scrubbed from every tool output before they reach the model.

---

Install

pip install exara-agent

Or with pipx for global isolation (recommended):

pipx install exara-agent

One-liner installers (detect Python, install pipx if missing, run wizard):

# macOS / Linux
curl -sSL https://raw.githubusercontent.com/santirivera-oss/exara-agent/main/scripts/install.sh | bash

# Windows (PowerShell)
iwr https://raw.githubusercontent.com/santirivera-oss/exara-agent/main/scripts/install.ps1 | iex

First run

exara init

Interactive wizard. Picks a provider, asks for an API key, optionally enables MCP servers (filesystem / github / memory / sequential-thinking) and writes a starter skill at ~/.ai-agent/skills/my-preferences.md.

Sanity check

exara doctor       # provider reachable? tools loaded? memory writable?
exara skills list  # which skills are active in this cwd
exara mcp list     # which MCP servers + tools are wired up

---

Use it

# One-shot task
exara run "summarise the structure of this project"

# Interactive REPL
exara chat

# Web UI (after `npm install` in frontend/)
exara serve   # http://127.0.0.1:8765
cd frontend && npm run dev   # http://localhost:3100

Slash commands

/help     /tools    /sessions       /resume <id>
/model    /plan     /init           /compact
/stats    /todos    /ps             /clear

---

Providers

Anything OpenAI-compatible works out of the box. Built-in presets:

Preset	Endpoint	Default model
ollama-local	http://localhost:11434	qwen2.5:7b-instruct
openrouter	https://openrouter.ai/api/v1	deepseek/deepseek-chat
openai	https://api.openai.com/v1	gpt-4o-mini
anthropic	https://api.anthropic.com/v1	claude-haiku-4-5
deepseek	https://api.deepseek.com/v1	deepseek-chat
groq	https://api.groq.com/openai/v1	llama-3.3-70b-versatile
lm-studio	http://localhost:1234/v1	(whatever you loaded)

exara profile presets       # show the catalog
exara profile add --from openrouter --use
exara profile use ollama-local

---

Permission levels

Level	Reads	Writes / shell / python	Confirmation
safe	yes	denied	n/a
normal	yes	allowed	required on high-risk
full	yes	allowed	skipped

Destructive command patterns (rm -rf /, mkfs.*, format c:, etc.) are blocked at every level via the bundled default_config.yaml.

---

Config layout

~/.ai-agent/
├── profiles.json        # provider keys (gitignored by design)
├── mcp.json             # global MCP servers
├── config.yaml          # optional global overrides
├── skills/              # global skills (always loaded)
│   └── my-preferences.md
├── data/agent.db        # conversation memory (SQLite)
└── logs/                # JSON logs

./                        # per-workspace overrides
├── .env                 # env vars (highest precedence)
├── config.yaml          # workspace-only overrides
├── mcp.json             # workspace MCP servers (merged on top of global)
└── skills/              # workspace skills (override globals by name)

Precedence (low → high): bundled defaults → ~/.ai-agent/config.yaml → workspace config.yaml → active profile → AI_AGENT_* env vars.

---

Architecture & docs

• docs/ARCHITECTURE.md — internals, ReAct loop, tool registry
• docs/MCP.md — how MCP integration works
• docs/ROADMAP.md — what's next
• CHANGELOG.md — release notes

---

Develop

git clone https://github.com/santirivera-oss/exara-agent
cd exara-agent
python -m venv .venv
. .venv/Scripts/Activate.ps1     # PowerShell on Windows
# source .venv/bin/activate      # macOS / Linux
pip install -e ".[dev]"
pytest                           # backend tests

# Web UI
cd frontend
npm install
npm run dev                      # http://localhost:3100

---

License

MIT — see LICENSE.