OSagent is an open-source AI agent with its own model and server built in. Plug into external APIs, add custom tools, or run entirely offline. Nothing leaves your machine unless you choose it.
Three reasons OSagent is built for your hardware, your data, your rules.
Tap to flip →
Own your stack
Load open weightsand tools on the machine you already pay for. No per-query tax to a gatekeeper — you buy silicon once, not a subscription to a black box API. Apache 2.0 code; the agent stays alive without rent.
Tap to go back →
Tap to flip →
Your machine, your boundary
The default path is on your hardware. Prompts and code don’t have to leave to “someone else’s data centre” first. Optional cloud is explicit, off by default — the trust boundary is the box in front of you.
Tap to go back →
Tap to flip →
Memory that doesn’t get wiped
Personalise without catastrophic forgetting. A durable toolgraph and Forget-Me-Not-style continuity — add skills and local models without erasing what the agent already knew.
Tap to go back →
OSagent is an open-source agent framework for local agents on your own machine. Capabilities on the left, a runtime demo on the right.
Five invocations, slash-style. Open to see the whole toolbelt.
/bashShell and scripts in your project.
/readLoad files into context.
/writePatches and edits, not wall-of-text pastes.
/searchRipgrep-style workspace search.
/httpHTTP to localhost and allow-listed URLs only.
Native command execution, same as your environment.
The agent can run the same bash and tooling you use — build, test, package — with output fed back for the next turn.
Model and tools run without a constant uplink.
Air-gapped and bad-network friendly for the core path. You decide when the agent reaches out.
Inspect, fork, ship — the stack is yours.
Apache 2.0 code, open weights, telemetry off. Add your own tools beside the five defaults — the framework stays small; your plugins live alongside.
Pick your local weights, cloud route, and skills — each group opens on click. The tables list the full set; the last column is a dropdown of every option so you can set the active choice from any row (all menus stay in sync).
Runtime stays on your machine. Choose a default load from the catalogue; every row’s Choose column lists the same full set.
| # | Model | Params | Quant | ~Size | Choose (all local) |
|---|---|---|---|---|---|
| 1 | Llama-3-70B-Instruct | 70.6B | Q4_K_M | ~41.5 GB | |
| 2 | DeepSeek-Coder-33B | 33B | F16 | ~66.0 GB | |
| 3 | Mistral-Nemo-12B | 12.2B | Q8_0 | ~13.1 GB | |
| 4 | Phi-3-Mini-128k | 3.8B | Q4_0 | ~2.2 GB | |
| 5 | Qwen2.5-72B | 72B | Q5_K_M | ~51.2 GB | |
| 6 | Gemma-2-9B-IT | 9B | Q6_K | ~6.1 GB |
Off-device route when you want it — or none to stay 100% local. Same pattern: one dropdown list repeated on every row.
| # | Route | Context / notes | Choose (all cloud) |
|---|---|---|---|
| 1 | None (local only) | No cloud calls | |
| 2 | Locai cloud | Pre-wired · same balance as app | |
| 3 | OpenAI API | Bring your key in settings | |
| 4 | Anthropic | Bring your key in settings | |
| 5 | Locai embeddings | Vector stage only | |
| 6 | Azure OpenAI | Enterprise route |
How big the pre-wired tool surface is — from the five default slash tools up to the full Forget-Me-Not graph.
| # | Pack | What you get | Choose (all skills) |
|---|---|---|---|
| 1 | Core (5 tools) | /bash · /read · /write · /search · /http | |
| 2 | Developer pack | Core + formatters + tests | |
| 3 | Research pack | Docs, fetch, summarize | |
| 4 | Forget-Me-Not (847) | Full toolgraph + memory | |
| 5 | Coding-heavy | Repos, LSP hints, patch apply | |
| 6 | Minimal | Only /read + /write |