tokensitterDelegated reads
The Tokensitter read and logs MCP tools delegate file and log reading to a cheaper model and return short answers with path-and-line excerpts. In the Codex plugin they use the user's own OpenAI key.
Features/Efficient file reading
Efficient file reading
Read what matters.
Spend less on the read.
Ask a cheaper model to read a file or log, then return a short answer with path-and-line excerpts. Use exact symbol reads when a bounded source slice is enough.
Add to Claude CodeOpenAI Codex: coming soon
Local MCP tools and read hooksEvidence snapshot: 27 Oct 2026 · per-lever live deployment not verified
Inside the workflow
Request a readA bounded file-reading task is identified.
Use a smaller helperA lower-cost model performs the eligible read.
Return useful contextThe main model receives the result for its task.
Transparent by designEvery claim needs a baseline
The technique
What changes.
And what does not.
Availability differs between the API relay and the local plugin. Whether each lever is deployed on the live relay has not been independently checked.
Exact symbol access
The Tokensitter outline, symbol and refs MCP tools expose symbols, exact numbered source and references for TypeScript and JavaScript, Python, Go and Rust. This is a bracket-aware outliner, not a compiler, and it refuses minified, huge or unbalanced files.
A guarded whole-file read
The local guard summarizes a whole-file read of a structured file with more than 300 lines, or directs other large reads toward bounded reads and helpers. Read once avoids sending an unchanged file again within its safety conditions.
The boundary. No dedicated measured saving is available for delegated file reading or symbol reads. The read-only helper test measures a related but different path; it is not a benchmark of every reading tool.
Quality checks
Fewer tokens is only
half the question.
Symbol extraction is bracket-aware, not a full compiler. There is no dedicated end-to-end quality or savings result for these reading tools, and the helper-cost study is not relabelled as file-reader evidence.
Checks that would strengthen this evidence. They are not yet reported for this feature:
Source fidelity
Compare claims, symbols, paths and cited lines against the actual files.
Relevant omissions
Check whether the helper dropped details required to solve the parent task.
Code outcome
Run the downstream build or tests, including tasks spanning multiple files.
Fallback behavior
Include long, ambiguous and dependency-heavy files that may require the main model.
How we compare
Same task.
A stated baseline.
No dedicated paired benchmark exists for these tools. The forced-Haiku helper experiment tests a different path and has its own page.
Value of the equivalent main-model read − value of the full helper read
We aim to report alongside: read-task correctness · net usage value per read · downstream task pass rate
We aim to report alongside: read-task correctness · net usage value per read · downstream task pass rate
Figures and evidence
The result.
With its boundaries.
Results are copied from our lab reports (snapshot 27 Oct 2026) and were not re-run for this website.
Source evidenceExact slicesSymbol tools return numbered source lines
Source evidenceNo model callSymbol outline and exact symbol reads
Source evidenceNot measuredDedicated delegated-read savings benchmark
Read this alongside the figures. No file-reader savings number is published. No claim is made that symbol reads or delegated reads match whole-file reasoning.
| Experiment | Workload and date | Result | Limits | Report |
|---|---|---|---|---|
| Implementation evidence | Described from the shipped plugin source · snapshot 27 Oct 2026 | Delegated reads, symbol tools, the large-read guard and Read once exist in the plugin. | No dedicated savings or quality result for these tools. |
|
Lab results at Anthropic API list prices on fixed tasks show what a lever did on those tasks, not what it will do on your bill. For plugin levers on Pro or Max, list-price values describe usage economics, not the subscription invoice. Relay levers apply to API keys only.
Read the measurement method and evidence limitsQuestions, answered
Does the helper benchmark prove file-reader savings?
No. It measures forced Haiku helpers on easy lookup tasks. Delegated read tools and symbol reads need their own evaluation. See Cheaper helpers for that study.
Are exact symbol reads summaries?
No. The symbol tool returns exact numbered lines for a selected symbol; the outliner refuses minified, huge or unbalanced files.
Keep exploring
A little more care.
A clearer picture.
Add Tokensitter to Claude Code and see where your tokens go.
Add to Claude CodeOpenAI Codex: coming soon