Table of Contents
Kestrel is a local-first research assistant that self-schedules and executes support tasks over long horizons so you can stay focused on key, complex & novel tasks. It handles the mechanical work—searching, filtering, extracting, and organizing evidence in the background, preparing it for review and analysis.
Designed for small local models. Kestrel is built to run with small LLMs (~1–7B) via Ollama, so it’s laptop/desktop-friendly (macOS, Windows, Linux; CPU-only works, GPU optional). No server cluster required.
- Task Exploration:: Leverages gathered evidence to iterate on tasks, identifying and pursuing new branches and next steps.
- Source Provenance: Uses a deep research approach, but maintains a clear record of consulted searches and how they impact exploration and findings.
- Long-running support: Keep tasks moving in the background and surface auditable checkpoints for review and redirection, rather than a single final report.
Example: Example: As you interpret a frontier paper and plan a novel study, Kestrel runs in the background to resolve the tooling (Polars vs. pandas vs. Spark), gathering real-world usage, extracting cited performance claims, and assembling a reviewable comparison.
- Research Agent — Executes concrete micro-tasks (search → retrieve → filter → extract → narrow summary) and proposes specific follow-ups when evidence warrants it.
- Orchestrator — Plans/schedules work, tracks progress and blockers, pivots when stalled, and emits checkpoints you can accept or adjust. This keeps efforts coherent over hours or days.
| Profile | Scope & behavior | Typical runtime |
|---|---|---|
| Hummingbird | Quick breadth-first sweeps; targeted lookups. | Minutes–~2h |
| Kestrel | Medium-horizon exploration with closely related branches. | Multi-hour |
| Albatross | Long-horizon, open-ended exploration with reviews. | Day-scale+ |
(Names are mnemonic. Community-authored profiles with clear behaviors are welcome—e.g., a Parrot profile for multi-agent dialogue.)
Screenshots
Home page |
Adding task |
- Backend: FastAPI
- Frontend: React
- Middleware: Redis
- Model Runner: Ollama
More model engines and integrations are under development!
Redis Queues and Channels:
REST API:
Ensure the following are installed:
- Ollama if not using the Dockerized version
- Recommended for Apple Silicon due to Lack of GPU Passthrough
- Docker Desktop
Clone the repository from Github:
git clone git@github.com:dankeg/KestrelAI.git
Ensure that Docker Desktop is running (easy way to check is running docker info in the terminal)
Navigate to the root of the repo, and run the following command to build and launch the application.
docker compose up --build
Alternatively, rather than building locally images are built through Github Actions, and can be fetched with docker pull first.
After a few minutes, the application will finish building. By default, it launches at http://localhost:5173/. Navigate here in a browser, and it's ready to use!
Using Kestrel is fairly straightforward. Create a new task, define a description for the research agents to base their exploration off of, and provide a time-box. Some examples are provided of tasks Kestrel can perform.
From there, the dashboard lets you monitor progress, such as tracking searches & the resulting sources, viewing checkpoints & reports, and exporting results.
See the open issues for a list of proposed features (and known issues).
- Top Feature Requests (Add your votes using the 👍 reaction)
- Top Bugs (Add your votes using the 👍 reaction)
- Newest Bugs
Reach out to the maintainer at one of the following places:
- GitHub Discussions
- Contact options listed on this GitHub profile
If you want to say thank you or/and support active development of Kestrel:
- Add a GitHub Star to the project.
- Tweet about the Kestrel.
- Write interesting articles about the project on Dev.to, Medium or your personal blog.
Together, we can make Kestrel better!
All contributions are greatly appreciated. Kestrel is meant to be a practical, community-focused project, focusing on supporting real-world use-cases. New feature suggestions are welcome, especially new orchestrator profiles tailored towards particular tasks of requirements, such as observability or structure.
Please read our contribution guidelines, and thank you for being involved!
The original setup of this repository is by Ganesh Danke.
For a full list of all authors and contributors, see the contributors page.
This project is licensed under the MIT license.
See LICENSE for more information.