relio 0.1.3

Relio — app-first AI framework with a built-in, governed memory engine

Relio

An app-first AI framework. Build a normal FastAPI + React app with your own data, and call AI in where you need it — memory, retrieval, agents, document extraction, and MCP — as one in-process component, not a pile of services.

Relio is memory-native: vectors + graph + structured query live in one SQLite file (or Postgres), and the AI reaches your app's data only through a governed, field-limited exposure map.

• One seamless system — backend calls the AI in-process (no network hop), one port, one deploy.
• AI is a component, not the product — the RelioAI seam is LLM-optional; use just recall, or the whole stack (chat, agents, extraction, MCP).
• Governed by default — the AI sees only what you declare; agents are bounded contexts with their own memory + tools.
• Batteries for building — scaffolds (web/mobile/desktop), generated SDKs, and a dev harness that won't let code land without a test and a doc.

Status: early / pre-release (0.1.x). The engine, server, SDKs, agents, extraction, and CLI are tested. Postgres paths and the vision/extraction model call are integration-gated (need a live DB / API). See Status.

---

Install

Requires Python 3.11+ (and Node 18+ only if you scaffold a web/mobile/desktop client).

pip install "relio[server]"      # engine + FastAPI server + Claude provider

Optional extras (combine, e.g. .[server,postgres]):

Extra	Adds
local	local ONNX embeddings (fastembed) — zero-API-cost vectors
mcp	the MCP server
postgres	Postgres + pgvector backend (with connection pooling)
server	FastAPI, uvicorn, Anthropic SDK
ai	the full AI-app stack for AIApp (server + local + mcp)
jwt	JWT auth hook (JWTAuth, via PyJWT)
dev	everything + pytest/coverage

Tip: use a virtual environment (python -m venv .venv) so the relio command lands on your PATH and deps stay isolated.

From source

git clone https://github.com/teamerisingstars/relio.git
cd relio
pip install -e ".[dev]"
pytest            # 170+ tests

---

Quickstart (use the AI component)

from relio import RelioAI

ai = RelioAI(path="relio.db")          # local SQLite + local embeddings, no LLM needed
ai.remember("Alice manages the Acme account")
print(ai.recall("who manages Acme?")[0].content)   # -> "Alice manages the Acme account"

Add an LLM when you want chat/extraction (set ANTHROPIC_API_KEY):

from relio import RelioAI
from relio.server.llm.claude import ClaudeProvider

ai = RelioAI(path="relio.db", provider=ClaudeProvider())
for chunk in ai.chat("what do you know about Acme?"):
    print(chunk, end="")

Scaffold a full app

relio new myapp --web      # FastAPI backend + React (Vite) frontend + generated SDK
cd myapp
relio dev                  # backend + Vite dev server on one URL

relio new also supports --mobile (Expo) and --desktop (Tauri), all on the same generated TypeScript SDK.

---

Core concepts

The RelioAI seam — the called-in component

One object composing the AI-system components. The LLM is optional.

ai.remember(text, scope=...)        # store          ai.recall(query)        # semantic retrieval
ai.embed(["a", "b"])                # batch embeddings ai.query(type=..., where=...)  # structured filter
ai.add_node / add_edge / neighbors / traverse          # knowledge graph
ai.chat(message)                    # agent loop (needs a provider)
ai.extract / ai.extract_file        # structured / multimodal extraction
ai.mcp_server()                     # expose to external agents over MCP

Exposure map — governed access to your data

Your app DB is private. The AI can call only what you declare, and see only the fields you allow.

@ai.tool
def lookup_account(name: str) -> dict:
    row = db.get_account(name)
    return ai.expose(row, fields=["name", "owner", "status"])   # cost/PII stay invisible

ai.call_tool("lookup_account", name="Acme")
# the same map auto-publishes as MCP tools: ai.mcp_server(include_tools=True)

Agents — bounded contexts

Each agent gets its own memory namespace, tool slice, config, and session. Private by default.

billing = ai.agent("billing", tools=["lookup_account"], system="You handle billing.")
billing.remember("invoice 42 overdue")     # not visible to other agents
billing.call_tool("refund", ...)           # PermissionError — not in its slice

Document extraction (AI beyond chat)

bom = ai.extract_file("drawing.pdf", schema={"properties": {"part_no": {}, "qty": {}}})

Run it as a server

from relio import Memory
from relio.server import create_app
from relio.server.llm.claude import ClaudeProvider

app = create_app(Memory(path="relio.db"), ClaudeProvider())   # uvicorn app:app
# memory-only backend (no LLM):  create_app(Memory())

Endpoints: POST /api/chat (SSE), /api/memory CRUD + /search + /query, /api/history, /api/graph/neighbors, /api/health. Identity comes from an auth hook (anonymous_auth default, or ApiKeyAuth / JWTAuth), never from the request body — so tenants are isolated by construction.

AI-first apps — AIApp (pip install "relio[ai]")

When the AI is the product, AIApp wires memory + bounded agents + a ready server in a few lines:

from relio import AIApp
from relio.server.llm.claude import ClaudeProvider

app = AIApp(provider=ClaudeProvider())
app.agent("assistant", system="You are helpful.")
asgi = app.build()   # adds GET /api/agents and POST /api/agents/{name}/chat (SSE)

Scaffold one with relio ai new <name>.

---

CLI

Command	Does
relio new <name> [--web/--mobile/--desktop]	scaffold an app (+ generated SDK + dev harness)
relio ai new <name>	scaffold an AI-first app (AIApp + a starter agent)
relio dev	run backend + frontend dev servers on one URL
relio build	build the React frontend
relio serve [--port]	serve API + built frontend on one port
relio sdk [--out]	generate the TS + Python SDKs from the API
relio develop ["<task>"]	drive the Claude Code CLI to build a feature (feeds gate gaps to it)
relio test [--coverage --min N]	run the test suites (optionally enforce coverage)
relio check	governance gate — fail if any module lacks a test and a doc
relio dockerfile / relio deploy	production Dockerfile / build image

The governance gate

A scaffolded app ships a CLAUDE.md (conventions), docs/, tests/, and a .claude/ Stop hook that runs relio check. The gate requires every module (Python and TypeScript) to have a test and a doc — and a fresh scaffold passes it out of the box. So agentic development with relio develop can't end with undocumented or untested code.

---

Backends & config

• Default: one SQLite file (vectors via sqlite-vec, WAL, indexed structured query). Backups = copy the file.
• Scale: Memory(database_url="postgresql://…") swaps to Postgres + pgvector (JSONB + GIN, connection pooling) — no caller changes.
• Server config via RELIO_* env (RELIO_DATABASE_URL, RELIO_MODEL, …).

---

Security

Secure-by-default (identity from an auth hook, never the request body; tenants isolated by scope; a governed, field-limited exposure map). Opt-in hardening in one call:

create_app(memory, provider,
    auth=ApiKeyAuth(keys, hashed=True),    # or JWTAuth(secret) — pip install "relio[jwt]"
    rate_limit=(60, 60),                   # 60 req / 60s per client
    max_body_bytes=1_000_000,              # 413 over the limit
    cors_origins=["https://yourapp.com"])

Destructive tools require @ai.tool(destructive=True) + confirm=True; validate model output with ai.extract(..., validate=True); SSE errors never leak internals. Full model, threat notes, and a deploy checklist in SECURITY.md.

---

Project layout

relio/            # the framework package
  ai.py           # RelioAI seam            exposure.py  # exposure map
  aiapp/          # AIApp (AI-first apps)   agents.py    # bounded agents
  memory.py       # the engine             backends/    # sqlite, postgres
  embedding/      # local + cache + batch
  server/         # FastAPI app, routes, auth (API-key/JWT), security, llm, agent loop
  cli/            # new/ai new/dev/serve/sdk/develop/test/check
  templates/      # web (React), mobile (Expo), desktop (Tauri)
docs/superpowers/specs/   # design specs (architecture v2 + every feature)
tests/            # 170+ tests

Start with docs/superpowers/specs/2026-06-30-relio-architecture-v2-app-first.md for the full architecture.

---

Status

Area	State
Engine, server, auth (API-key/JWT), graph, query, agents, AIApp, security, SDK gen, scaffolds, CLI, dev harness	✅ tested
Postgres + pgvector (backend, pooling, JSONB)	⚙️ implemented; integration-gated (RELIO_TEST_DATABASE_URL, pytest -m integration)
Claude vision/extraction call	⚙️ implemented; untested without an API key (the offline fake path is tested)
Generated TS SDK / mobile / desktop apps	⚙️ scaffolded + structurally tested; not compiled in CI

---

License

Intended open-core: the framework is MIT (see LICENSE); a future managed cloud + enterprise components are separately licensed. See the project's licensing notes.

Contributing

See CONTRIBUTING.md. In short: tests-first, every change keeps the suite green, and design specs live under docs/superpowers/specs/.