Repository navigation
MiniAgent v1.0.0
MiniAgent is a lightweight, readable Python agent harness for studying and
testing the boundaries between a model, an authorization layer, Tools, state,
observability, and MCP.
Highlights
- Sync and native async Agent Runtime
- Provider-independent model interface with deterministic
FakeModel - OpenAI-compatible Qwen and DeepSeek support
- Ordered single and multi-Tool Calling with ID continuation
- Context, explicit Memory, and JSON checkpoint/recovery
- Trace, metrics, evaluation, and failure attribution
- Tool risk metadata, policy, confirmation, retry, and idempotency
- Official MCP v2 discovery/invocation with allowlisting and namespacing
- Optional local NVIDIA GPU diagnostics with fake and Windows backends
The normal installation and all 208 offline tests require no API key, GPU,
CUDA Toolkit, or PyTorch. Install from the repository with:
python -m pip install .
python examples/basic_agent.pyThe real DeepSeek -> MiniAgent -> MCP -> NVIDIA GeForce GTX 1650 integration
was verified with 3 model calls, 5 successful GPU Tool Calls, 0 Tool failures,
8,440 total tokens, and Evaluation PASS. A later audit confirmed policy and
confirmation already occurred before MCP invocation; the follow-up correction
changed Trace event placement only, and no post-audit paid rerun is claimed.
Safety
Model output requests actions; it never grants itself authority. Every Tool
Call passes schema validation, trusted metadata, ExecutionPolicy, optional
confirmation, and retry/idempotency handling before execution. MCP discovery
is not authorization. MiniAgent exposes no arbitrary shell, PowerShell, Python
execution, or filesystem Tool.
Major limitations
Tool batches are sequential, provider output is non-streaming, the GPU backend
targets device 0, JSON persistence is unencrypted, idempotency is process-local,
and checkpointing cannot provide exactly-once external side effects. GPU
thresholds are diagnostic heuristics, and live Qwen/DeepSeek availability
depends on the external provider.
MiniAgent is released under the MIT License. See README.md, SECURITY.md,
and RELEASE_CHECKLIST.md for usage, security, and publication details.