Mingbird gets small 2–9B local models to finish real multi-step tasks on a laptop, running offline against a local Ollama backend.
Yes, Mingbird is open source under the Apache-2.0 license.
Mingbird is free to use.
Yes, Mingbird can be self-hosted (the source is available under the Apache-2.0 license).
Mingbird is a local-first agent harness that gets small 2–9B models to finish real multi-step tasks on an ordinary laptop, running offline against a local Ollama backend. Open source: yes (Apache-2.0); self-hostable with a one-click zero-outbound mode and all config in ~/.ollama_agent/config.json; free OSS. It is brand-new (~1 star, created 2026-09-17) but active — 30 commits, v1.8.2, a 461-test suite, and CI on Linux and macOS.
Mingbird is the scaffolding around a small local model, not another model. Its bet is that most agent failures are harness defects — bad tool wiring, lost state, no retries — rather than the model being too small, so it focuses on the loop that turns a 2–9B model running under Ollama into something that actually completes a multi-step task on a normal laptop. You run it fully offline: it talks to a local Ollama backend, offers a one-click zero-outbound mode so nothing leaves the machine, and keeps its configuration in ~/.ollama_agent/config.json. The project ships a 461-test suite and a 288-cell public benchmark of harness-vs-model behavior. Open source: yes (Apache-2.0); self-hostable; free OSS.
Most GTM agent work does not need a frontier model: classify a reply, draft a follow-up, extract fields, decide the next step. Mingbird is a bet that a small model on your own laptop can do that reliably if the harness around it is good — which, if it holds, means an offline, zero-cost, zero-data-egress agent for the routine steps of an outbound or ops workflow. The honest read: this is a brand-new project (~1 star) and the "it's the harness, not the model" thesis is the maker's own, benchmarked only by the maker. It belongs on the watch list as part of the small-local-model lineage (alongside jeff and jeeves for decisions, openshell for isolation) — try it on a low-stakes task first and judge completion quality on your own data before it runs anything that touches a real system.