Browser agent: local models (Ollama + browser-use)

Paperful’s browser_agent lane and paperful recover drive browser-use against your session vault Chromium profile. The model must emit tool-shaped actions (click by element index, wait, navigate) — not chat about downloading.

There is no official browser-use benchmark for local Ollama tags. The guidance below is evidence-weighted from browser-use docs, GitHub issues, and community reports (2025–2026), folded for Paperful’s honesty rules and shipped wiring. See ROADMAP § browser-use.

How Paperful uses browser-use (not vanilla quickstart)

Topic	Paperful behaviour
LLM	[browser_agent].model if set, else [llm].model; Ollama → ChatOllama, LiteLLM → ChatLiteLLM. Optional fallback_model: one full retry on miss (not on captcha / bot wall). Tags with / (e.g. openai/gpt-4o-mini) use LiteLLM on the fallback attempt and need llm.allow_remote when provider = "ollama".
Vision	[browser_agent].use_vision = false by default — DOM / element index only. Set true only with a VL-capable tag (qwen2.5vl, etc.); doctor / recover preflight reject text-only tags when vision is on.
Browser	System Chrome (channel="chrome"), same profile as session login; uBlock + cookie extensions on
Success	Valid PDF on disk (min_pdf_bytes, stable size) — not agent done alone
Task	Stay on DOI/landing; no search engines; no purchase; stop on CAPTCHA / paywall / 403
Caps	[browser_agent].max_steps, max_wall_s; agent stopped early when PDF lands
Batch recover	paperful recover --from-last-run replays keys from state/last-run.json (--from-last-run-mode, --limit). Same runner as --item.
Learned playbooks	Agent PDFs append state/fetch-wins.jsonl with promotable click: / rewrite wins and optional steps trace → playbooks propose / promote.

Operator setup: LLM § recover, sessions, config § browser_agent. Install via uv sync --extra browser-agent or a heavy Compose image (Docker); headed vault login stays on the host.

Recommended Ollama tags (by VRAM band)

Use ollama pull tags that support tool calling (ollama show <tag> — avoid HTTP 400 “does not support tools”). Prefer non-thinking instruct tags for recover; thinking models can burn step/wall time.

VRAM (rough)	Good defaults	Notes
24 GB+	qwen2.5vl:32b (hard UI + vision), qwen3:32b / qwen3:30b, qwen2.5:32b	Best local band for publisher cookie → PDF clicks. For VL tags set use_vision = true.
12–16 GB	qwen3:14b, qwen2.5:14b	Documented floor for recover; more retries, tight prompts. Match use_vision=False.
~8 GB	Prototype only: llama3.1:8b, qwen2.5vl:7b	Official docs’ Ollama example is llama3.1:8b — connects, often fails soft walls.

Paperful doctor ambers agent models whose name looks under ~10B (e.g. 7b in the tag). llm.md recommends 14B+ for browsing. Treat 14B as the serious minimum; 32B-class for batch recover on messy publisher UIs.

Ranked starting points (PDF / publisher UI)

1. qwen2.5vl:32b — vision + layout; best when controls are icon-only or missing from the a11y tree — set [browser_agent].use_vision = true and this tag as model (or fallback_model).

2. qwen3:32b / qwen3:30b — strong tool/agent loops (Ollama tool docs use Qwen3).

3. qwen2.5:32b — proven text JSON / multi-step UI in community reports.

4. mistral-small:24b — function-calling specialist; fits mid VRAM.

5. qwen3:14b / qwen2.5:14b — minimum serious band.

6. qwen2.5vl:7b — small vision option.

7. llama3.1:8b — docs plumbing example only; not a production PDF agent.

Upstream warns that non–qwen-vl-max Qwen can emit flat action JSON (e.g. {"navigate": "url"} instead of nested objects). browser-use docs suggest concrete nested examples in the system prompt if parsing fails.

Avoid as primary recover models

• Sub-7B chat tags, embedding models, legacy qwen: (pre-2.5) without tools.

• Many Gemma3 builds → Ollama 400 (no tools template) in community reports.

• DeepSeek-R1 / heavy reasoning tags — thinking noise, slow steps, fragile tools API.

• Small coder-only tags — JSON sometimes works; browsing/UI loops are weaker than general agent models (validate on your publishers before relying on a coder tag).

• Hosted-only names (qwen-vl-max, Browser Use Cloud bu-*) — not ollama pull.

Context, temperature, and Ollama tuning

browser-use sends large DOM snapshots each step. Ollama may default to a small context window on consumer GPUs (community reports cite ~4k under 24 GiB). Agent workloads often want 16k–32k+; 64k when VRAM allows. Paperful does not yet pass num_ctx into ChatOllama — raise context in Modelfile / Ollama env or track ROADMAP for a config knob.

Community configs often use temperature=0 for action selection. browser-use maintainers note that smaller local models struggle with tool-calling (discussion #4261).

Acceptance test (one page, your rights)

Before trusting a tag for a collection run:

1. Pick one article URL you already may access (OA or campus session in the vault).

2. paperful recover --item <KEY> --dry-run then run without dry-run (or recover --from-last-run --dry-run after a run that hit the agent).

3. Pass = PDF under out/ with sane bytes — not a polite done in logs.

4. Fail = invents a new host, opens a search engine, or claims success with no file (Playwright download caveat).

Paperful already enforces disk verification and aborts on search/support URLs; models that only succeed by hallucinating PDF URLs should be demoted regardless of size.

Deterministic paths beat the agent

Use HTTP / vault Playwright lanes first. Reserve browser_agent for soft UI (cookie banner → “Download PDF”) when metadata lanes and vault retry already failed. See architecture § sources and workflows.

LiteLLM / remote

With provider = "litellm", page text may leave the machine (orange disclaimer). Use for recovery only with allow_remote understood. Reject ollama/… ids on the LiteLLM path — use provider = "ollama" instead. See llm.md.

References

• browser-use: supported models, tools, agent params

• Ollama: tool calling

• Issues: #442, #486, #814

• Internal research note (2026-09-29): distilled into this page; full multi-lane draft kept under assessments/2026-09-29-ollama-browser-use-pdf-download-models.md.