MCP Tools DirectoryAbout this integration
Traces which neurons and attention heads drive a language model's output via MCP tools.
- Transport
- stdio
- Authentication
- none
- Initial setup
- none
- Runtime
- unattended
- Evidence
- documented
- Version
- 0.1.5
- Package
- neuronscope-cli
- Last compatibility test
- Not independently tested
Connect your agent
neuronscope-cliExisting identity and connection fields are preserved. Publisher documentation was reviewed without installing dependencies, downloading a model, or running inference. The first public-model call downloads weights; untrusted use needs external CPU and memory limits, and circuit results are approximate single-component analyses rather than full path patching.
Capabilities: Attention-head and neuron attribution tracing, Layer activation summaries, Single-component zero-ablation, Approximate circuit sketches
Connected profiles
Additional details
io.github.RudrenduPaul/neuronscope
Source ↗ · Checked 2026-09-170.1.5
Source ↗ · Checked 2026-09-17CC0-1.0; package licenses are separate
Source ↗ · Checked 2026-09-17pypi
Source ↗ · Checked 2026-09-17neuronscope-cli
Source ↗ · Checked 2026-09-190.1.6
Source ↗ · Checked 2026-09-19The MCP client spawns the local Python server over stdio; model loads and forward passes have documented concurrency and size caps but no built-in wall-clock timeout.
Source ↗ · Checked 2026-09-19The documented public-model setup requires no credential and explicitly notes the expected unauthenticated Hugging Face Hub rate-limit warning; gated models may have separate upstream access requirements outside this baseline connection.
Source ↗ · Checked 2026-09-19The existing listing remains stdio and the reviewed configuration launches neuronscope-mcp through uvx from neuronscope-cli.
Source ↗ · Checked 2026-09-19Publisher documentation reviewed; package and integration endpoint not executed or independently security-audited.
Source ↗ · Checked 2026-09-19