> ## Documentation Index
> Fetch the complete documentation index at: https://docs.agentguardian.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Scan an Ollama-backed agent

> Red-team an agent whose model backend is a local Ollama instance.

Ollama runs open-weights models locally on `http://localhost:11434`. The
target AgentGuardian scans is still an HTTP agent — the agent's *model
backend* happens to be Ollama. Point `--endpoint` at the agent.

## What this example tests

* All 10 ASI categories against an HTTP-exposed agent backed by an
  on-device model.
* The `ollama:` model spec for the swarm itself (`--model
  ollama:llama3.1`) — graded scans with no external API costs.

## Prerequisites

* AgentGuardian installed — `pip install agent-guardian`.
* Ollama installed and running locally: `ollama serve`.
* A model pulled into Ollama: `ollama pull llama3.1`.
* The bundled
  [`examples/ollama_local/`](https://github.com/glacien-technologies/agent-guardian/tree/main/examples/ollama_local)
  fixture, or your own agent that calls Ollama.

## Run target

Host install:

```bash theme={null}
ollama serve &
ollama pull llama3.1
uv sync --extra examples
uv run uvicorn examples.ollama_local.serve:app --port 8000
```

Or Docker Compose (Ollama + the FastAPI agent in one stack):

```bash theme={null}
docker compose -f examples/ollama_local/docker-compose.yml up
docker compose -f examples/ollama_local/docker-compose.yml exec ollama ollama pull llama3.1
```

The bundled agent reads `OLLAMA_BASE_URL` and `AG_DEMO_OLLAMA_MODEL` from
the environment so both invocation patterns work without code edits.

## Run AgentGuardian

```bash theme={null}
agent-guardian scan \
  --endpoint http://localhost:8000/chat \
  --model stub \
  --mode fast \
  --output md \
  --output-path scan.md
```

For a graded run with the same Ollama instance driving the swarm:

```bash theme={null}
agent-guardian scan \
  --endpoint http://localhost:8000/chat \
  --model ollama:llama3.1 \
  --mode smart \
  --output md \
  --output-path scan.md
```

Flag-by-flag, every option below is verified against
`src/agent_guardian/cli.py`:

* `--endpoint URL` — hosted HTTP endpoint of the agent.
* `--model stub` / `--model ollama:<name>` — stub is offline; the Ollama
  spec drives the swarm with the same local instance the agent uses.
* `--mode fast` — `fast` / `smart` / `full` (default).
* `--output md --output-path scan.md` — Markdown report.

## Expected output

`examples/ollama_local/sample-scan.json` ships a committed reference
scan for `--model stub`. The `notes` field documents what changes when
you swap in `ollama:llama3.1`.

## Common errors

* **`ollama-demo: cannot reach Ollama at http://localhost:11434`.** The
  bundled agent returns this string when Ollama is not running. Start
  it with `ollama serve` and re-pull the model.
* **Scan stalls with `--model ollama:<name>`.** The swarm is blocked
  waiting for Ollama to load the model into memory. Run a single warm-up
  prompt against Ollama first (`curl -s http://localhost:11434/api/chat
  -d '{"model":"llama3.1","stream":false,"messages":[{"role":"user","content":"hi"}]}'`).
* **`Pull complete` but scan still uses a different model.** Ollama
  serves the model named in the request — confirm
  `AG_DEMO_OLLAMA_MODEL` matches the model you pulled.

## Next step

* For a full local-CI loop with no external dependencies, combine the
  Ollama stack with [GitHub Actions](/ci-cd/github-actions) using a
  self-hosted runner.
* For a tool-bearing target, read
  [Scan a LangGraph agent](/try/scan-langgraph).