paperprism-agent 0.2.2

Local-first Agent for PaperPrism: receives arxiv download events from the Chrome extension and mirrors papers into a hidden vault with LLM-powered tagging.

PaperPrism Agent

Local HTTP service that receives archive events from the PaperPrism Chrome extension, mirrors arxiv PDFs into a hidden workspace vault (~/.paperprism/vault), enriches them with arxiv-API + PDF metadata, classifies them with the user's LLM, generates TL;DR summaries, auto-tags every new paper, and exposes a small REST API that the extension's Dashboard / Options / Topic pages talk to.

Stack

• Python 3.10+
• FastAPI + uvicorn
• Pydantic v2
• SQLite (with FTS5) for metadata, tags, topics, and jobs
• PyPDF + arxiv API for enrichment
• Pluggable LLM backends: OpenAI / Anthropic / Google Gemini / Qwen / DeepSeek / Moonshot / OpenRouter / Ollama (configured via ~/.paperprism/llm.yaml + secrets.env)

Quick start

The Agent ships on PyPI as paperprism-agent. Pick whichever install mode suits you.

Option A — uvx (zero-install, one-off)

Best for "just try it" or CI smoke tests. No venv to manage.

# Needs uv 0.4+ — see https://docs.astral.sh/uv/
uvx paperprism-agent serve

Each invocation runs in a throwaway environment, so this is not suitable for paperprism-agent install (launchd needs a stable path — see Option B).

Option B — uv tool install (recommended for daily use)

Installs a stable shim at ~/.local/bin/paperprism-agent that launchd can call into. Upgrades are uv tool upgrade.

uv tool install paperprism-agent
paperprism-agent serve            # foreground test
paperprism-agent install          # register launchd LaunchAgent (macOS)

Option C — pipx / pip in a venv

pipx install paperprism-agent
# or
python3 -m venv .venv && source .venv/bin/activate
pip install paperprism-agent
paperprism-agent serve

Option D — editable checkout (for contributors)

cd agent
python3 -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'
paperprism-agent serve
# or equivalently:
python -m paperprism_agent serve

Default bind: http://127.0.0.1:17321.

Smoke test

# Health
curl -s http://127.0.0.1:17321/api/health | jq .

# Fake an archive.completed event pointing at any local PDF
curl -s -X POST http://127.0.0.1:17321/api/ingest \
  -H 'Content-Type: application/json' \
  -d '{
    "event": "archive.completed",
    "arxivId": {"id":"2401.08281","fullId":"2401.08281v1","version":"v1","legacy":false},
    "sourceUrl": "https://arxiv.org/pdf/2401.08281v1.pdf",
    "absUrl": "https://arxiv.org/abs/2401.08281",
    "downloadPath": "/absolute/path/to/any.pdf",
    "triggerClassification": true,
    "emittedAt": "2026-04-27T10:00:00Z"
  }' | jq .

You should see the file mirrored under:

~/.paperprism/vault/YYYY/MM/2401.08281v1/
├── paper.pdf
└── meta.json

…and, if an LLM is configured and auto_tag_on_ingest is on (default), the paper will gain 2–5 tags visible via GET /api/tags within a few seconds.

HTTP API (summary)

All endpoints are rooted at http://127.0.0.1:17321. Auth header X-PaperPrism-Token required only if --token (or PAPERPRISM_TOKEN) is set.

Method	Path	Purpose
GET	/api/health	Liveness + version
POST	/api/ingest	Extension's archive.requested / archive.completed event sink
POST	/api/ingest/upload	Multipart PDF upload used by the Dashboard bulk-import
GET	/api/papers	Paginated paper list with filters (tag, topic, free text)
GET	/api/papers/{id}	Single paper detail (metadata + tags + topics)
DELETE	/api/papers/{id}	Remove a paper (and its vault files)
GET	/api/papers/{id}/pdf	Stream the archived PDF
GET	/api/dimensions/values	Distinct values for each LLM dimension (for Dashboard filters)
GET	/api/tasks/stats	Per-dimension counts (for Dashboard headline stats)
GET	/api/tags	List all tags + counts
GET	/api/papers/{id}/tags	Tags attached to a given paper
POST	/api/papers/{id}/tags	Edit tags: body {"add":[...],"remove":[...]}
POST	/api/tags/auto	Start a batch auto-tag + topic-synthesis job for the given paper ids
GET	/api/tags/auto/{job_id}	Poll status of an auto-tag job
DELETE	/api/tags/auto/{job_id}	Cancel a running auto-tag job
POST	/api/tags/auto/{job_id}/retry	Retry the failed papers of a finished job
GET	/api/topics	List topics (each with name, summary, all tags)
GET	/api/topics/{slug}	Topic detail + papers
DELETE	/api/topics/{topic_id}	Delete a topic (papers keep their tags)
GET	/api/llm/config	Read current LLM provider / model / toggles
PUT	/api/llm/config	Update LLM config (written to llm.yaml + secrets.env)
POST	/api/llm/test	Tiny chat request to verify the configured key

Concrete request/response schemas live in paperprism_agent/models.py and paperprism_agent/server.py, and the TypeScript client is extension/lib/agent.ts.

CLI

Command	Purpose
paperprism-agent serve	Run HTTP server in the foreground
paperprism-agent install	macOS: install launchd LaunchAgent and start it
paperprism-agent uninstall	macOS: stop and remove LaunchAgent
paperprism-agent status	macOS: print launchctl state
paperprism-agent restart	macOS: force launchd to (re)start the service
paperprism-agent logs	Tail logs (--which out|err|launchd-out|launchd-err, --follow)
paperprism-agent version	Print version

Flags for serve / install:

• --host, --port, --token, --home
• serve also: --log-level
• Env fallbacks: PAPERPRISM_HOST, PAPERPRISM_PORT, PAPERPRISM_TOKEN, PAPERPRISM_HOME

Background autostart on macOS

paperprism-agent install     # writes ~/Library/LaunchAgents/com.paperprism.agent.plist
                             # and bootstraps it into gui/<uid>
paperprism-agent status      # shows launchctl state
paperprism-agent logs --follow
paperprism-agent uninstall   # stop + delete plist

The LaunchAgent:

• starts at user login (RunAtLoad)
• auto-restarts on crash (KeepAlive.Crashed), not on clean stops, so uninstall and graceful SIGTERM actually work
• rate-limits respawns (ThrottleInterval=10) to avoid hot-looping
• captures launchd-level stdout/stderr to logs/launchd.{out,err}.log while the app's structured log continues to flow into logs/agent.{out,err}.log
• uses sys.executable from the venv you ran install from, so upgrading is just pip install -e . && paperprism-agent restart

Linux (future)

A systemd --user unit will be generated by install when the platform is Linux. For now, run paperprism-agent serve in a terminal or wire it into your own systemd user unit.

Filesystem layout

~/.paperprism/
├── runtime.json        # {port, pid, token, version}
├── llm.yaml            # provider, model, api_base, api_key_env, auto_tag_on_ingest, ...
├── secrets.env         # API keys (mode 600)
├── db.sqlite           # papers, tags, paper_tags, topics, jobs, FTS5
├── logs/
│   ├── agent.out.log
│   ├── agent.err.log
│   ├── launchd.out.log
│   └── launchd.err.log
└── vault/
    └── YYYY/MM/<arxivId>/
        ├── paper.pdf
        └── meta.json

db.sqlite is managed by numbered migrations under paperprism_agent/migrations/ and applied automatically on startup whenever the file's schema_version is behind the code.

LLM configuration

~/.paperprism/llm.yaml (also editable from the extension's Options page → LLM section) shape:

provider: qwen            # one of: openai, anthropic, google, qwen,
                          # deepseek, moonshot, openrouter, ollama
model: qwen-plus
api_base: https://dashscope.aliyuncs.com/compatible-mode/v1
api_key_env: QWEN_API_KEY  # name of the env var holding the key
enrichment_enabled: true   # pull arxiv API metadata + PDF abstract
classification_enabled: true
auto_tag_on_ingest: true   # LLM-tag every paper added via /api/ingest
pdf_full_char_limit: 8000 # max chars of PDF full text fed to LLM classifier

Secrets are never written into llm.yaml; they sit in ~/.paperprism/secrets.env at mode 600 and are loaded into the Agent's process environment on startup (and live-injected when the Options wizard saves a new key).

Contract with the extension

See extension/lib/agent.ts. The server validates incoming JSON with paperprism_agent.models. Breaking changes on either side must bump meta.schema_version and ship a matching SQLite migration.