Deep Research agent designed to work fully on-premise. It uses local models so no data leaves your local network. It can use tools to browse the internet if given permission to do so.
To use it, see install instruction below, then just go to a folder you want to work on and type
droplet
this will activate the command line agent which will summarize the content of the current folder and given you some options to start doing deep research
Below the BrowseCompPlus [1] baseline scores for the DROP agent using gpt-oss-120b and two context compaction strategies: summarization and keep-last-N messages. Numbers are averages over three seeds.
| Experiment Group | Accuracy (%) |
|---|---|
| Drop-agent baseline | 65.5 (0.1) |
| recursive summarization (LLMR10) | 67.6 (0.4) |
| Keep-last-5 (KL5R20) | 68.5 (1.1) |
See scripts/bcp for evaluation details.
Install via pip (needs at least Python 3.12), for example
uv venv -p 3.12
source .venv/bin/activate
uv pip install git+ssh://git@github.com/IBM/drop-agent.git
If you want to use Ollama as your backend (good option for local laptop usage), install the Ollama server, in OSX
brew install ollama
To start Ollama now and restart at login:
brew services start ollama
if you have to launch vLLM yourself, you would run on $HOSTNAME
vllm serve $MODEL --host 0.0.0.0 --port $PORT
If you have access to a vLLM server, run droplet with the vLLM backend:
droplet -b vllm -m $MODEL -u http://${HOSTNAME}:${PORT}
you can also use e.g. --save-config remote-vllm to store this config (with that port) for later use, with
droplet -c remote-vllm
any further arguments will override the defaults above
Clone and install in editable mode (here uv is used, but pip works too):
git clone git@github.ibm.com:generative-computing/drop-agent.git
cd drop-agent
uv venv -p 3.12
source .venv/bin/activate
uv pip install --editable .