Neuronscope

NeuronScope

[CI] [PyPI] [License: MIT]

[NeuronScope - Traces LLM outputs to the neurons and heads that caused them | Product Hunt]

Ask a language model "why did you say that" and get back the actual attention heads and neurons responsible, as JSON, from the command line or from an agent over MCP.

[NeuronScope tracing a real gpt2 prediction from the command line, showing the top attention heads and MLP neurons responsible for the output]

Install

pip install neuronscope-cli

That gets you the neuronscope command. To install from source instead (for development or to track main):

git clone https://github.com/RudrenduPaul/NeuronScopecd NeuronScopepip install -e .

[!NOTE] The first run of any command downloads the requested model from the HuggingFace Hub (gpt2 is about 500MB) and prints two lines to stderr that are expected, not errors: a CPU-fallback notice if you don't have a CUDA GPU, and an unauthenticated-HF-Hub rate-limit notice. Neither one means anything broke.

Quickstart

neuronscope trace gpt2 "The capital of France is Paris. The capital of Japan is" --top-k 5

Real output from this exact command (stderr trimmed to the two expected warnings mentioned above):

Prompt: The capital of France is Paris. The capital of Japan isPredicted next token: ' Tokyo'Top attention heads (by direct logit            attribution)┏━━━━━━━┳━━━━━━┳━━━━━━━━━━━━━━━━━━━┓┃ Layer ┃ Head ┃ Logit attribution ┃┡━━━━━━━╇━━━━━━╇━━━━━━━━━━━━━━━━━━━┩│     9 │    8 │            4.0679 ││     8 │   11 │            2.9028 ││    10 │    7 │           -1.4782 ││     8 │   10 │           -1.3999 ││    10 │    0 │            1.1424 │└───────┴──────┴───────────────────┘Top MLP neurons (by activation          magnitude)┏━━━━━━━┳━━━━━━━━┳━━━━━━━━━━━━┓┃ Layer ┃ Neuron ┃ Activation ┃┡━━━━━━━╇━━━━━━━━╇━━━━━━━━━━━━┩│    10 │     97 │     7.8394 ││    11 │    611 │     4.6954 ││    11 │   2997 │     4.6468 ││    10 │   1793 │     4.5443 ││     9 │   1460 │     4.4196 │└───────┴────────┴────────────┘

gpt2 predicts Tokyo correctly, and head L9H8 is the single biggest contributor to that prediction. Add --json to get the machine-readable version of the same result:

neuronscope trace gpt2 "The capital of France is Paris. The capital of Japan is" --top-k 3 --json

{  "schema_version": 1,  "operation": "trace",  "model": {    "requested_name": "gpt2",    "resolved_name": "gpt2",    "backend": "transformer_lens",    "device": "cpu",    "n_layers": 12,    "n_heads": 12,    "d_model": 768,    "d_mlp": 3072  },  "prompt": "The capital of France is Paris. The capital of Japan is",  "predicted_token": " Tokyo",  "predicted_token_id": 11790,  "top_neurons": [    { "layer": 10, "neuron_index": 97, "activation": 7.839381217956543 },    { "layer": 11, "neuron_index": 611, "activation": 4.695372581481934 },    { "layer": 11, "neuron_index": 2997, "activation": 4.646785736083984 }  ],  "top_heads": [    { "layer": 9, "head_index": 8, "logit_attribution": 4.067923545837402 },    { "layer": 8, "head_index": 11, "logit_attribution": 2.9028172492980957 },    { "layer": 10, "head_index": 7, "logit_attribution": -1.4781968593597412 }  ]}

What it does

NeuronScope is a CLI and MCP server built on top of TransformerLens. TransformerLens does the actual model loading, hooking, and activation math; NeuronScope adds a stable CLI, a versioned JSON schema, and an MCP server around it, so a script or an agent can ask "which components drove this output" without writing TransformerLens code directly.

• trace: runs a prompt through the model and ranks attention heads by direct logit attribution to the predicted token, and MLP neurons by activation magnitude at the final prompt position.
• activations: dumps shape, mean, std, min/max, and the max-activating sequence position for every layer's residual stream, MLP neuron activations, and attention pattern.
• patch: zero-ablates one component (resid_pre, resid_mid, resid_post, attn_out, mlp_out, or mlp_post) at a given layer and reports how the predicted token and its logit changed.
• circuit: a best-effort automated circuit sketch. Ranks candidate heads/neurons by logit attribution, then measures each one's individual causal effect via single-component ablation. This is not full path-patching with clean/corrupted prompt pairs and does not capture interaction effects between components. The --json output says so explicitly in its method field.
• Every command supports --json for a schema_version-stamped document instead of a table, and the same four operations are exposed as MCP tools returning the identical shape via .model_dump(), so a CLI call and an MCP tool call produce the same document for the same input.
• Model support is whatever transformer_lens.HookedTransformer.from_pretrained supports. Installing neuronscope-cli today pulls TransformerLens 3.6.0, which supports 249 pretrained checkpoints and aliases (OFFICIAL_MODEL_NAMES), covering GPT-2, Pythia, Llama, Gemma, Qwen, and more. Small models like gpt2 run comfortably on CPU.

NeuronScope does not replace TransformerLens, nnsight, SAELens, Anthropic's circuit-tracer, or Neuronpedia. It wraps TransformerLens for one narrower job: fast, scriptable, agent-callable component tracing on a single prompt. It leaves deeper mechanistic work (SAE training, transcoder-based circuit graphs, hosted feature browsing) to those tools.

CLI reference

Every command takes MODEL (any name HookedTransformer.from_pretrained accepts, for example gpt2 or EleutherAI/pythia-70m) and PROMPT as positional arguments.

Global: neuronscope --version, neuronscope <command> --help. Exit codes: 0 success, 1 a runtime error (prompt too long for the model's context window, --layer out of range, etc.), 2 a Click usage error (bad flags), 3 an unsupported model name.

[neuronscope circuit ranking candidate heads/neurons by logit attribution and measuring each one's causal effect via single-component ablation]

[neuronscope patch zero-ablating one component at a given layer and reporting how the predicted token and its logit changed]

MCP Server

NeuronScope ships a Model Context Protocol server so an AI agent (Claude, Cursor, or any MCP-compatible client) can trace, inspect, ablate, and sketch circuits directly, without a human invoking the CLI by hand.

Install the extra:

pip install "neuronscope-cli[mcp]"

Add it to your MCP client's config (for Claude Desktop, claude_desktop_config.json):

{  "mcpServers": {    "neuronscope": {      "command": "uvx",      "args": ["--from", "neuronscope-cli", "neuronscope-mcp"]    }  }}

The server exposes four tools, trace, activations, patch, and circuit, each returning the identical pydantic-model-shaped JSON the CLI's --json flag prints, via .model_dump(), so an agent calling this server and a script calling the CLI get the same document for the same input. A real trace call and its response:

trace(model="gpt2", prompt="The capital of France is Paris. The capital of Japan is", top_k=3)
{  "schema_version": 1,  "operation": "trace",  "predicted_token": " Tokyo",  "predicted_token_id": 11790,  "top_neurons": [    { "layer": 10, "neuron_index": 97, "activation": 7.839381217956543 }  ],  "top_heads": [    { "layer": 9, "head_index": 8, "logit_attribution": 4.067923545837402 }  ]}

Errors never raise across the tool boundary: every handler catches its exceptions and returns a structured ErrorResponse dict instead, so a calling agent always gets a parseable result.

[!WARNING] NeuronScope caps model size (2B parameters by default, NEURONSCOPE_MAX_MODEL_PARAMS) and how many models can load at once (1 by default, NEURONSCOPE_MAX_CONCURRENT_LOADS), but puts no timeout on model loading or forward passes. If you expose this MCP server somewhere an untrusted agent can call it, still put a resource limit around the process (a cgroup, ulimit, or a container memory/CPU cap) as defense in depth rather than relying on these in-process caps alone.

Transport is stdio, so there is nothing to host: the MCP client spawns the server as a local subprocess. Source: neuronscope/mcp_server.py.

• Claude Code reads this from a project-level .mcp.json in your repo root, or you can add it with claude mcp add neuronscope -- neuronscope-mcp.
• Claude Desktop reads this from its claude_desktop_config.json (~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows), under the same "mcpServers" key.
• The neuronscope mcp-server CLI subcommand still works as a local, non-uvx alternative that runs the same server over stdio from an existing install.

How it compares

All five of these are real, actively maintained projects doing different jobs. This table compares CLI/JSON-agent-output surface and model coverage, not depth of interpretability research, where TransformerLens, nnsight, SAELens, circuit-tracer, and Neuronpedia are all more mature than NeuronScope. Star counts, release info, and last-push dates below were pulled from each project's GitHub API on 2026-08-03 and will drift over time; check the repos directly for current numbers.

The honest differentiation is narrow: NeuronScope is the only one of these with a CLI, a native MCP server, and a versioned JSON schema together in one package, and it's model-agnostic across whatever TransformerLens supports rather than pinned to a fixed transcoder allowlist like circuit-tracer. It is not more capable, more mature, or more widely used than any of these projects.

What is NeuronScope and why does it exist

TransformerLens gives you a Python API for loading a model and running hooked forward passes. That's the right interface for a research notebook. It's the wrong interface for a script that needs a subprocess call and a JSON document back, or for an agent that needs a tool it can call over MCP. NeuronScope exists to be that second interface: the same underlying computation, wrapped so a CLI invocation or an MCP tool call gets back a schema-versioned document instead of a Python object graph.

FAQ

Is this a replacement for TransformerLens, nnsight, SAELens, circuit-tracer, or Neuronpedia? No. NeuronScope is built directly on TransformerLens and does not do anything TransformerLens itself can't already do at a lower level. It doesn't train SAEs (SAELens), do full path-patching circuit discovery with transcoders (circuit-tracer), give you a Python-native tracing context manager for arbitrary PyTorch models (nnsight), or host a browsable feature database (Neuronpedia). It's a CLI and MCP wrapper around one slice of TransformerLens's functionality.

What models are supported? Anything transformer_lens.HookedTransformer.from_pretrained supports, which today is 249 checkpoints and aliases spanning GPT-2, Pythia, Llama, Gemma, Qwen, and others. Run python -c "from transformer_lens.loading_from_pretrained import OFFICIAL_MODEL_NAMES; print(len(OFFICIAL_MODEL_NAMES))" in your own environment to get the exact count for your installed version, since TransformerLens adds models over time.

Does it need a GPU? No. Small models like gpt2 run fine on CPU; that's what the test suite and the quickstart above run on. Larger models will be slow on CPU. NeuronScope does not auto-select Apple Silicon's MPS backend even when available, because PyTorch's MPS backend can silently produce incorrect values for some ops that this project's activation-patching math depends on being exact. Pass device="mps" explicitly in your own code if you want it anyway.

Is it safe to expose the MCP server to an untrusted agent? Only with resource limits in place. See Known limitations below.

How is NeuronScope different from circuit-tracer, the other CLI tool in this list? circuit-tracer does deeper circuit analysis (full attribution graphs from trained transcoders) but only for a fixed allowlist of models: Gemma-2, Gemma-3, Llama-3.1/3.2, Qwen-3, and GPT-OSS. NeuronScope trades that depth for breadth: it works with any of TransformerLens's 249 supported checkpoints with no transcoder training step, and ships an MCP server so an agent can call it directly. The cost is that NeuronScope does single-component logit attribution and zero-ablation, not transcoder-based path patching.

Does the installed version always match what's on PyPI? Run neuronscope --version after installing to check. pip install neuronscope-cli pulls whatever release PyPI has published most recently; the code on this repo's main branch can be ahead of that between releases. Installing from source (pip install -e .) always tracks main exactly, including whatever hasn't been released yet.

What license is NeuronScope under, and can I use it commercially? MIT. You can use, modify, and redistribute it in commercial and closed-source projects, with attribution and the license notice kept intact. The dependencies it pulls in (TransformerLens, PyTorch, the mcp package) carry their own licenses; check those separately if you're redistributing a bundled product rather than just calling neuronscope-cli as a dependency.

Known limitations

• circuit is an approximation. It ranks components by logit attribution and measures each one's individual causal effect via single-component zero-ablation on one prompt. It does not do full path-patching with clean/corrupted prompt pairs, and it will not catch interaction effects between components. The --json output states this in its method field so a caller doesn't have to trust prose to know the caveat.
• No timeout on model loading or forward passes. Once a request passes the resource caps below, NeuronScope runs the load and the forward pass to completion with no built-in wall-clock limit. If you run the MCP server somewhere an untrusted agent can call it, put a resource limit around the process (a cgroup, ulimit, or a container memory/CPU cap) as defense in depth.
• Model size and load concurrency are capped, but only in-process. neuronscope/core/limits.py rejects a model over NEURONSCOPE_MAX_MODEL_PARAMS (2B parameters by default) before any weights are downloaded, and rejects a load once NEURONSCOPE_MAX_CONCURRENT_LOADS (1 by default) other loads are already in flight, both with a structured error rather than a hang or a crash. The size check is best-effort: if a model's parameter count can't be determined (for example, fully offline with nothing cached yet), it fails open rather than blocking a legitimate request, so it's not a hard guarantee on its own -- pair it with a process-level resource limit for untrusted deployments.
• HookedTransformer.from_pretrained is deprecated upstream. TransformerLens 3.6.0 emits a DeprecationWarning pointing at TransformerBridge.boot_transformers as the replacement. It still works today, and every command shown in this README ran on it, but NeuronScope's backend hasn't migrated yet. Tracked as an open item; migrating would be a change inside neuronscope/backends/transformer_lens.py, not a change to any CLI command or MCP tool signature.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md for dev setup, where the code lives, and what a PR needs before it merges. Quick version:

pip install -e ".[dev,mcp]"pytest -v

CI runs the same suite on Python 3.10, 3.11, and 3.12 on every push and pull request against main. The suite covers 87% of neuronscope/ (pytest --cov=neuronscope), with the MCP server's less-exercised paths (specific error branches) the main gap.

License

MIT. See LICENSE.