K-Dense BYOK (bring your own keys) is a free, open-source AI research assistant for scientists in any field that runs on the researcher's own computer. The researcher supplies access to a model of their choice, hosted or running locally, and the application supplies everything else: a place for the work to run, a layer of scientific scaffolding, and a complete record. Each project is an ordinary folder, so the data, the code, the results, and the record stay on a machine the researcher administers and can be read years later without the application. Three things separate it from a chat assistant or a general-purpose coding agent. It ships a library of written scientific procedures, guided workflow templates, catalogs of where research data can be found, and reviewer and writer roles the agent can hand work to. It keeps a Living Lab Notebook whose entries link into an argument and are added to but never erased. And it records what happened by watching what the agent does rather than by taking the agent's word for it, in a log the agent has no tool that can write to. That choice targets the most common failure, model overclaiming, in our earlier benchmark of nine frontier models, by making claims checkable rather than preventing them. On twenty interdisciplinary research prompts, scored under a rubric fixed in advance, K-Dense BYOK led two managed platforms on both scientific quality and research execution. Its deliverables were the only ones that recorded the software they ran in, and the only ones that usually arrived with a command that regenerates the results. One of the managed platforms ran the same frontier model and supplied neither. Those environment records were files the agent wrote, not part of the observed log, which does not yet capture the software environment itself. The code is available under the MIT license at https://github.com/K-Dense-AI/k-dense-byok.
Figures & tables
Figure 1: How K-Dense BYOK works, from the researcher’s side. The researcher brings a question, data, and access to a model; Kady, the agent, does the work in a project folder on the researcher’s own machine, drawing on the scaffolding shown at the top, and returns ordinary files. Two records are written as it works: Kady writes the Living Lab Notebook, and the application writes the provenance log by observing Kady’s actions. Because both refer to the same files, each file a notebook entry cites can be reported as current, stale, or unverified.
Layer
Organization
Count
Scientific skills
procedures followed with ordinary tools
163
Guided workflow templates
across 22 categories
326
Data-resource entries
across 18 categories
229
Specialists
helper agents Kady can hand work to
21
Previewable scientific formats
structures, spectra, imaging, arrays
60+
Table 1: Scientific scaffolding shipped with K-Dense BYOK 0.9.14. The workflow, resource, and specialist counts were verified against the source repository. The skill catalog is copied into each project from the Scientific Agent Skills library the first time the project starts ( Kassis et al., 2026 ) ; that library is versioned separately and keeps growing, so the count given here is the one current at the time of writing, and the application’s own README still reports an earlier figure of 149.
Limit
Value
What happens when exceeded
Files scanned per comparison
20,000
step marked as too large to compare fully
Largest file fingerprinted
512 MB
size and time stamp kept; marked not fingerprinted
File links recorded per step
200
extra links dropped; the number dropped is reported
Tool arguments stored per step
4 KB
stored as a shortened preview
Comparison fails or has no baseline
–
step marked as failed or as lacking a baseline
Table 2: Limits in the provenance recorder and how they are reported. Each limit marks the affected step instead of truncating its record unannounced; the values were read from the source code.
Figure 2: What each platform’s deliverables contained, over the twenty prompts, for five of the seven items on the standard list. Cells show how often each platform’s bundles contained each item. These checks measure only what is needed for an audit and do not establish scientific correctness.
Scientific quality
Research execution
K-Dense BYOK
86.0 [82.2, 89.9]
64.5 [60.6, 68.2]
Claude Science
74.8 [68.7, 80.0]
26.0 [22.7, 29.7]
Biomni Lab
64.9 [56.1, 72.7]
38.3 [33.4, 42.4]
Difference vs. Claude Science
+11.2 [6.2, 17.2]
+38.5 [33.2, 43.5]
prompts led by K-Dense BYOK
16/20
20/20
Difference vs. Biomni Lab
+21.1 [13.2, 30.1]
+26.2 [21.9, 30.9]
Table 3: Judged outcomes over the twenty prompts, as % of available points (scientific quality out of 75, research execution out of 25), with 95% bootstrap confidence intervals obtained by resampling prompts. The lower rows give the paired difference by prompt, in percentage points, between K-Dense BYOK and each comparison platform, and the number of prompts on which K-Dense BYOK scored higher than that platform. Intervals are recomputed from the per-prompt scores in Table 4 (Section 7 ).
Figure 3: Summary of platform performance. (A) Mean outcomes with 95% bootstrap confidence intervals over prompts. (B) Distributions of the secondary 0–100 combined score, with diamonds marking platform means. (C) Number of first-place finishes on each outcome; tied platforms share first place.
Figure 4: How often each platform placed first, second, or third on the two outcomes. Bars count prompts; ranks are assigned within each prompt, and tied platforms share the better rank, so a lower rank may go unused. On research execution K-Dense BYOK ranked first on all twenty prompts and neither comparison platform ranked first on any.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Scientific quality
Research execution
Combined
Prompt
KD
CS
BL
KD
CS
BL
KD
CS
BL
1
82
70
68
62
24
50
77
59
64
2
88
34
19
62
12
8
82
28
16
3
68
74
66
47
17
37
63
60
58
4
84
73
61
70
24
37
80
61
55
5
82
72
69
70
24
37
79
60
61
Appendix
Table 4: Scores of all sixty bundles, as % of available points for scientific quality (out of 75) and research execution (out of 25) and as points out of 100 for the secondary combined score, rounded to whole numbers. Prompts are numbered as in Appendix B . KD, K-Dense BYOK; CS, Claude Science; BL, Biomni Lab.
Figure 5: The two outcomes plotted against each other, without combining them into one score. Small points are results on individual prompts; large markers with whiskers are platform means with 95% bootstrap confidence intervals. K-Dense BYOK separates from both comparison platforms mainly along the research-execution axis, while Claude Science and Biomni Lab swap places between the two axes.
Figure 6: Both outcomes across the twenty benchmark prompts. Large markers show platform means with 95% bootstrap confidence intervals over prompts; faint points are the scores on individual prompts. Left panel: scientific quality as % of 75 available points. Right panel: research execution as % of 25 available points. The spread between platforms is visibly wider on research execution than on scientific quality.
Figure 7: Scores for each of the twenty benchmark prompts. Each row is one prompt and each column one platform; cells give scientific quality and research execution as percentages of available points, and the secondary combined score in points out of 100, all on one shared 0–100 color scale that is used for display only. The research-execution panel separates the platforms on nearly every prompt, while in the scientific-quality panel the platforms trade places on several prompts.
Figure 8: Mean rating on each criterion over the twenty prompts. Cells show mean judge ratings on the 0–4 scale, without weighting, so the criteria can be compared directly. Scientific-quality criteria (A–I) are at left and research-execution criteria (R1–R4) at right. K-Dense BYOK leads on all nine quality criteria but by modest margins. The main separation is concentrated on executable regeneration (R1) and environment and determinism (R3).
ID
Criterion
Points
A
Task fulfillment
13
B
Evidence and data integrity
11
C
Methods and validation
13
D
Result correctness and reasoning
11
E
Robustness and uncertainty
6
F
Decision quality and follow-up
6
Appendix
Table 13
Score
Anchor
4
Complete, correct, specific, and auditable. All central requirements are met and no material defect is found.
3
Substantially correct and complete. Only minor omissions or defects that do not change the conclusion.
2
Partially correct. A material omission, weak choice, or error reduces confidence, but useful work remains.
1
Minimal or seriously flawed. Major requirements are absent, unsupported, or incorrect.
0
Absent, nonresponsive, contradicted by evidence, fabricated, or fundamentally wrong.