Agno Demo

A minimal AgentOS to copy and build on. A small set of agents and a local SQLite database for sessions + memory.

For a production-ready version of this demo, see the agent-platform-railway codebase. Comes with AgentOS (FastAPI) + Postgres. One-command deploy to Railway. JWT auth, Slack integration, eval suite, and recursive-improvement loops driven by Claude Code.

Agents

Agent	What it does	Backing
LocalWiki	Read + write a local markdown wiki. Ingest URLs via the web — "add a page about X" fetches, digests, and files in one update call.	`WikiContextProvider(FileSystemBackend, web=ParallelMCPBackend)`
GitWiki (env-gated)	Same as LocalWiki, but the wiki lives in a real git repo. Auto-commits and pushes after each write. Registered when `WIKI_REPO_URL` + `WIKI_GITHUB_TOKEN` are set.	`WikiContextProvider(GitBackend, web=ParallelMCPBackend)`
NotionWiki (env-gated)	Same as LocalWiki, but the wiki is a Notion database (one row per page). Writes round-trip through Notion blocks; the database is the source of truth. Registered when `NOTION_API_KEY` + `NOTION_DATABASE_ID` are set.	`WikiContextProvider(NotionDatabaseBackend, web=ParallelMCPBackend)`
WebSearch	Keyless web research via Parallel MCP. Returns answers with cited URLs.	`WebContextProvider(ParallelMCPBackend)`
CodeSearch	Answers questions about this repository — file paths, line numbers.	`WorkspaceContextProvider`
Researcher	Composes WebSearch + LocalWiki + CodeSearch on one agent. Checks the wiki first, searches the web, queries the codebase, and files findings back into the wiki.	composition of the three providers above

All agents share db=get_db() (SQLite at data/demo.db), agentic memory on, datetime + history in context, markdown output.

Teams

Team	What it does	Pattern
Swarm	Broadcast the same question to two web-search agents — one on OpenAI gpt-5.5, one on Anthropic claude-opus-4-7 — and synthesize. Lead calls out where the models agree, disagree, and a confidence read.	`Team(mode=broadcast, members=[openai_agent, anthropic_agent])`

This is the "assemble a bunch of agents on a common problem, mix OpenAI and Anthropic" pattern. Both members share the same web provider — one Parallel MCP session, two perspectives.

Get started

1. Create a virtual environment

bash

uv venv .venvs/demo --python 3.12
source .venvs/demo/bin/activate

2. Install dependencies

bash

uv pip install -r cookbook/01_demo/requirements.txt

3. Set your API keys

bash

export OPENAI_API_KEY="..."
export ANTHROPIC_API_KEY="..."   # required for the Swarm team's Claude member

export PARALLEL_API_KEY="..."    # optional — raises rate ceiling on Parallel MCP

# Optional — enables the GitWiki agent
export WIKI_REPO_URL="https://github.com/<owner>/<repo>.git"
export WIKI_GITHUB_TOKEN="ghp_..."   # PAT with contents:write

# Optional — enables the NotionWiki agent
export NOTION_API_KEY="ntn_..."          # integration token
export NOTION_DATABASE_ID="..."          # UUID from the database URL

4. Serve

bash

fastapi dev cookbook/01_demo/run.py

Then open os.agno.com and sign in:

Add OS → Local
Connect to http://localhost:8000, call it Local AgentOS
Chat with your agents

Evals

From the repo root:

bash

python -m cookbook.01_demo.evals               # run all cases (concise)
python -m cookbook.01_demo.evals -v            # stream the full agent run
python -m cookbook.01_demo.evals --case <name> # run one case

Or from cookbook/01_demo:

bash

python -m evals
python -m evals -v
python -m evals --case <name>

Each case runs the agent (or team / workflow) once, then checks the response with AgentAsJudgeEval (LLM rubric, binary pass/fail) and optionally ReliabilityEval (tool-call assertion). Results log to SQLite — connect AgentOS at os.agno.com to see history.

Extending

To add an agent: drop a file in agents/, register it in run.py's AgentOS(agents=[...]) list, add quick prompts to config.yaml, restart. Same pattern for teams/ and workflows/. Add eval cases in evals/cases.py once it's stable.

Regenerating requirements

bash

./cookbook/01_demo/generate_requirements.sh

Edits to requirements.in are the source of truth; the .txt is regenerated and pinned via uv pip compile.