Get This Tool
Subtext
Pricing
- Model
- Free
Summary
You set a breakpoint, you read the logits, you stare at the output — and you still have no idea what the model was 'thinking' between prompt and token. Subtext exists for that gap.
Subtext is an open-source, self-hosted tool that surfaces a local language model's internal representations — specifically the J-space global workspace identified in Anthropic research — as the model reads and generates, before output tokens appear. You get a browser-based live view and the ability to record, export, and replay sessions for later analysis. The stack is Python-based and runs against local models, so there is no cloud dependency and no data leaves the machine. The ceiling appears quickly: Subtext has no API, no integration hooks, and no support for models it cannot instrument directly — which means anything running behind a remote endpoint stays opaque.
Bottom line: Pick this if you are a researcher or developer running local models and want to see internal states that logs will never show you — but plan a different approach the moment your model lives on a remote server or you need to pipe observations into a broader evaluation pipeline.
Community Performance Report Card
No community ratings yet. Be the first to rate this tool!
Community Benchmarks Community
Sign in to submit a benchmarkNo community benchmarks yet. Be the first to share a real-world data point.
Pros
Sign in to edit- Exposes J-space internal representations in real time before output tokens are committed, so you can observe the model's reasoning process rather than reverse-engineering it from outputs after the fact.
- Session recording and browser-based replay via `record_session.py`, which means you can share a captured session with a colleague or step through a failure case hours after it happened without reconstructing the exact prompt conditions.
- Fully self-hosted with no cloud dependency, so sensitive prompts and model states never leave the machine — critical for any research involving proprietary data or unpublished model weights.
- Ships with a `verify_accuracy.py` script and a linked reference paper, so teams can audit whether the J-space instrumentation is behaving as the methodology describes rather than taking the visualization on faith.
Cons
Sign in to edit- No API and no integration hooks mean every observation stays inside the tool. Teams who want to log J-space states to a database, trigger alerts on specific internal patterns, or feed observations into an evaluation framework have to parse exported files themselves — that DIY layer adds maintenance burden fast.
- Instrumentation is limited to local models the tool can attach to directly. The moment your model moves to a hosted endpoint — OpenAI, Anthropic, or any other remote API — Subtext goes dark. Teams running production models in the cloud will find nothing here to replace their existing black-box debugging approach, and at that point they switch to tools built around token-level logging or external evaluation harnesses instead.
- Multi-turn session tooling exists (`test_multiturn.py`) but the project has 25 commits and no open issues or pull requests — community surface area is thin. Teams who hit an edge case with a specific model architecture have no forum, no community plugin layer, and no documented escalation path beyond reading the source.
Community Reviews
Sign in to write a reviewNo reviews yet. Be the first to share your experience.
About
- Platforms
- Python, local inference
- API Available
- No
- Self-Hosted
- Yes
- Last Updated
- 2026-07-14T14:08:44.611Z
Best For
Who it's for
- Researchers studying LLM internals
- Developers debugging local model behavior
- Users wanting to visualize J-space representations
What it does well
- Observing model reasoning during live chat
- Inspecting internal states before output tokens appear
- Exporting and replaying model sessions for analysis
Discussion Community
Sign in to commentNo discussion yet. Sign in to start the conversation.
Spotted incorrect or missing data? Join our community of contributors.
Sign Up to ContributeCommunity Notes & Tips Community
Sign in to contributeBe the first to contribute. General notes, observations, gotchas, and tips from people who use this tool day-to-day.
Frequently Asked Questions
- Is Subtext free?
- Yes — Subtext is fully free to use. There is no paid tier.
- Is Subtext open source?
- Yes. Subtext is open source.
- Can I self-host Subtext?
- Yes. Subtext supports self-hosting on your own infrastructure.
- What platforms does Subtext support?
- Subtext is available on: Python, local inference.
Hours Saved & ROI Stories Community
Sign in to contributeBe the first to contribute. Concrete time/cost savings, with context. e.g. "Cut my code review backlog from 4h to 45m per week."
Curated lists that include this category
Most debugging tools for language models show you what came in and what came out. Subtext shows you what happened in between. The tool attaches to a locally-running language model and streams its J-space representations — internal states the model uses for multi-step reasoning — to a browser-based dashboard in real time. You can watch the verbal workspace shift as the model reads the prompt and before it commits to output tokens. Sessions can be recorded via `record_session.py`, exported, and replayed in the browser, so you can step through a conversation after the fact and inspect state at each turn.
The differentiating feature is the J-space instrumentation itself. Grounded in Anthropic research identifying a small set of internal representations that behave like a global workspace — contents that the model can verbally report, deliberately modulate, and causally use for reasoning — Subtext makes those representations observable rather than inferred. That is a different class of visibility than attention weights or token probabilities, and the repo includes a reference implementation and a paper link for teams who want to validate or extend the methodology.
Subtext fits tightly into one workflow: a researcher or developer with a local model who wants ground-truth visibility into reasoning steps, not just output quality. It breaks equally fast in two directions: any model running behind a remote API is outside its reach, and there are no hooks to feed observations into an external evaluation framework, a database, or a CI pipeline. Teams doing one-off debugging sessions will find it self-contained. Teams who need to track internal states across hundreds of runs and correlate them with downstream metrics will hit the export-and-do-it-yourself wall immediately.
The project ships with `server.py`, `test_client.py`, `test_multiturn.py`, and `verify_accuracy.py` alongside a `requirements.txt`, with Windows batch start support via `start.bat`. The interface is a single `index.html` served locally. There is no packaged binary and no install wizard — you clone, install dependencies, and run the server against a compatible local model.