A WorkBuddy Comparison Limited to Local Traces and Official Differences
A WorkBuddy Comparison built from this machine is a record of local traces and official permission differences, not a benchmark. There was no side-by-side test, so no verdict on speed, quality, or which tool is better is published. What it holds is a full archive for one tool and thinner evidence for three others. Values were checked on 2026-11-10.
01 /What exactly does this page compare?
This WorkBuddy Comparison is deliberately narrow. It compares the shapes of local records left by four tools on one machine, and it sets those beside the official permission and sandbox rules. It does not compare features, pricing, or output quality. The reason is simple: only the first two kinds of information were actually observed or published.
Four names appear in the local evidence: WorkBuddy, Codex, Cursor, and Claude Code. Their configurations and partial traces coexist in this user's home directory, which is the entire scope of the observation. Nothing here was collected on another machine or from a controlled run. A single machine on a single date is a small sample, and the page treats it as one.
In this WorkBuddy Comparison the word comparison is used in a literal sense. Two sets of facts are placed next to each other: what the directories contain, and what the official documentation says about permissions. No third set, such as timing or accuracy, was gathered. Any sentence that would require such a measurement is left out.
02 /What did this machine actually leave behind?
The WorkBuddy Comparison evidence begins with the largest local archive. Its projects directory holds 87 session directories, 2505 files, 481 MB, and 256 .jsonl files, with sessions and tool outputs intact. A separate traces directory holds a further 1371 files. That is the only tool on this machine with a complete session archive.
Codex leaves a different shape. The logs_2.sqlite logs table contains 1000 rows, and the computer-use folder holds Codex Computer Use.app (26.831.1000926). But goals_1.sqlite and state_5.sqlite report 0 rows for thread goals, threads, and projects, so no session thread records remain on this machine. config.toml contains only a desktop section, one MCP server section, and nine plugin sections.
Cursor leaves workspace traces without a session archive. Its projects directory holds 38 directories, 1512 files, and 6.9 MB, including real project names and browser MCP tool definitions, and a skills-cursor folder holds 28 skills. The mcp.json file names only lark-openapi-mcp and github, and the agents folder is empty. Claude Code leaves the thinnest record of all: settings.json is the only file in that directory, with hooks and statusLine, and no projects, history, or sessions; that is a statement about this path, not a claim that the tool was never used.
For the official description of the product, What Is WorkBuddy, a plain bounded description is the starting point. It is background reading rather than evidence for a verdict. The WorkBuddy Answers index is the parent of this page. Neither link supplies a measurement that would change the boundary stated here.
| Tool | What is present in the local record | What is missing from it |
|---|---|---|
| WorkBuddy | 87 session directories, 2505 files, 481 MB, 256 .jsonl files, plus 1371 files in a traces directory | Nothing is missing from the session archive |
| Codex | 1000 rows in the logs_2.sqlite logs table; Codex Computer Use.app (26.831.1000926) in the computer-use folder | 0 rows for thread goals, threads, and projects; no session threads |
| Cursor | 38 project directories, 1512 files, 6.9 MB, 28 skills in skills-cursor | mcp.json lists only lark-openapi-mcp and github; the agents folder is empty |
| Claude Code | settings.json in the inspected directory, with hooks and statusLine | No projects, history, or sessions at this path |
03 /What do the official permission and sandbox rules say?
The WorkBuddy Comparison also draws on official documentation, checked on 2026-11-10. Default permission is bounded by the workspace, and "allow full access" is a global switch in the current client that does not carry across devices. Batch-delete approval defaults to 50 items, and the automatic backup limit defaults to 3000 MB. Security gateway, end-to-end encryption, and sensitive-data protection are on by default, while the three AI-safety switches are off by default.
The cloud sandbox has its own limits. It does not support workspace access, full access, or custom models, and it runs in an isolated Tencent Cloud sandbox without needing the computer to stay online. In both modes, AI inference happens in the cloud while local task scripts run on the machine, and conversation credits are counted against the task owner. Scheduled tasks draw on three sources, local, project, and cloud, and connector reads and writes consume no credits while an expert team consumes several times what a single expert consumes.
Non-official Skills carry a separate set of official warnings. They may contain malicious prompt injection, privilege escalation, or backdoors, and they may send input to third-party services. When a Skill calls a third party, modifies local files, or executes system commands, it does so as you, and the scripts/ directory runs in the client sandbox with an official disclaimer about the results. Complex tasks and complex Skills usually consume more credits, but the official Skills Market documentation has not published a specific per-call credits number, checked on 2026-11-10.
| Official setting | Documented default or limit |
|---|---|
| Default permission | Bounded by the workspace |
| "Allow full access" | A global switch in the current client, not carried across devices |
| Batch-delete approval | 50 items |
| Automatic backup limit | 3000 MB |
| Security gateway, end-to-end encryption, sensitive-data protection | On by default |
| Three AI-safety switches | Off by default |
| Cloud sandbox | Does not support workspace, full access, or custom models |
| AI inference in both modes | Runs in the cloud; local task scripts run on the machine |
04 /Which differences are official and which are only local records?
The WorkBuddy Comparison separates two kinds of claim. Official defaults come from documentation and apply to the product generally, while directory counts come from one machine on one date. A number such as 256 .jsonl files describes an archive, not a capability. Confusing the two would turn a local observation into a product claim, which this page does not do.
The same separation governs the other tools. A logs table with 1000 rows is a fact about a database file, not evidence that a session thread existed. An empty agents folder is a fact about a folder, not evidence about the editor's behaviour. Only the shape of the record is being compared, never the tool behind it.
Two further cautions come from the local evidence file. The word "Cursor" inside project directories is often the technical term for a pointer rather than a reference to the editor, so it is not counted as usage evidence. The MEMORY.md file is a user self-report, not a runtime log. It records the MCP layer as two tiers tested on 2026-09-25, GSC wrapped as a global Skill, and a source_mode column in the sessions table.
| Boundary | Status on this page |
|---|---|
| WorkBuddy session archive | Full sessions and tool outputs are present |
| Codex logs | 1000 rows exist, but no session threads |
| Claude Code | No traces at this path, which is not the same as never used |
| Speed, output quality, which tool is better | Not published; no side-by-side test was run |
05 /Why is there no side-by-side test?
There was no side-by-side test on this machine. The WorkBuddy Comparison therefore publishes no conclusion about speed, no conclusion about output quality, and no conclusion about which tool is better. Those three verdicts are withheld because none of them was measured. A benchmark would require running the same task on each tool under the same conditions, and that did not happen.
The local evidence file states the same boundary in plain terms. Configurations and partial traces for WorkBuddy, Codex, and Cursor coexist here, but only one tool left a complete session archive. Codex has logs without session threads, Cursor has workspace traces, and Claude Code has no traces at this path. The site therefore writes only provable boundaries and the shape of the local records, not a ranking.
That refusal, in this WorkBuddy Comparison, is not a hedge. A claim that one tool is faster would need timings, and a claim that one is better would need an agreed task and a scoring rule. Neither exists in the material used for this page. Leaving the verdict out is the only honest option.
06 /How should you read these numbers?
The WorkBuddy Comparison numbers are counts of files, rows, and directories, taken on 2026-11-10. They are not measurements of how the tools perform, and they do not generalise to another machine. A different user would almost certainly show a different set of traces. Treat the table as a snapshot of one home directory, not as a survey.
Counting rules matter in this WorkBuddy Comparison too. A session directory is counted once, and a .jsonl file is counted as a file, not as a conversation of a fixed length. Zero rows in a database table means no record was found there, not that the feature is absent from the product. Where the evidence supports only a statement about this path, the page says exactly that.
07 /What can you legitimately conclude?
The WorkBuddy Comparison supports a modest conclusion: one tool left a complete, dated session archive on this machine, and the others left thinner or no traces at the paths inspected. It also supports the official differences listed above, checked on 2026-11-10. It does not support any claim about which tool is preferable. Adoption decisions need evidence this page does not have.
That is the practical use of this WorkBuddy Comparison. If you are choosing a tool, the official permission and sandbox rules are the part you can verify from documentation, and the local traces are only a record of what one person left behind. Nothing here substitutes for running your own task on your own files. The page is a boundary statement, not a recommendation.
08 /Where should you go next?
If you need the product definition first, What Is WorkBuddy, a plain bounded description is the place to start. It is background rather than evidence. the WorkBuddy Answers index is the parent of this page. Both links are navigation, not measurements.
WorkBuddy Skills and their permission layers is the page to read before enabling anything that writes files. That is a permission question, not a performance one. About WorkBuddy Central is the site's trust page. Neither link changes the boundary stated above.
No other link is needed to use this WorkBuddy Comparison. The four local records and the official defaults are the whole of the evidence. Anything beyond them, including a verdict, is deliberately absent. The date at the top of this page is 2026-11-10.
If you arrived here from a search rather than from the site, the WorkBuddy guide home page is the better entry point: it states the scope and the limits first, and this page only makes sense underneath them.
09 /Frequently asked questions
What does the WorkBuddy Comparison actually cover?
It covers the shape of local records left by four tools and the official permission and sandbox differences, checked on 2026-11-10. It does not cover speed, output quality, or which tool is better. Those would need a side-by-side test that was not run. The page states that absence explicitly rather than filling it with a guess.
Why does only one tool have a complete session archive?
Because that is what this machine happened to record. The assistant left 87 session directories with 256 .jsonl files, while Codex kept logs but no session threads, Cursor kept workspace traces, and Claude Code left only a settings file at the inspected path. That is a statement about these directories, not about the tools in general. A different machine would show a different shape.
Do the official defaults apply to every mode?
No, the cloud sandbox has its own limits. It does not support workspace access, full access, or custom models, and AI inference runs in the cloud while local task scripts run on the machine. Workspace-bounded default permission and the global full-access switch describe the client, checked on 2026-11-10. The documentation separates those cases, and so does this page.
Will you publish a winner later?
Only if a real side-by-side test is run and its method can be shown. Until then, no conclusion on speed, quality, or which tool is better will appear here. A verdict without measurements would be a fabricated benchmark. The site would rather leave the comparison narrow than publish a confident guess.
Cost is deliberately outside this comparison. The published API token prices for DeepSeek V4.1 Flash, GPT-5.1 Codex and Claude Opus 5, together with the workload arithmetic and the two vendor pages I could not open from this machine, are recorded on a separate page about automation and per-token cost. A price table is not a benchmark, and neither page publishes a verdict on which tool is better.
10 /Sources
The WorkBuddy Comparison uses only two research files and the official differences recorded in them, read on 2026-11-10. Local counts come from a single inspection of this machine's home directories, checked on 2026-11-10. Official defaults come from the Automation Guide, Permission Modes, the cloud sandbox documentation, the Skills Market, and the non-official Skill security notice, checked on 2026-11-10. Figures relayed by media coverage of an enterprise security white paper are not reproduced here, because they are not official documentation.
WorkBuddy Central is an independent site. Not affiliated with Tencent, OpenAI, Anthropic, or any product mentioned.
WorkBuddy is a product of Tencent. Codex is a product of OpenAI. Claude Code is a product of Anthropic. Cursor, Trae, Qoder and CodeBuddy belong to their respective owners. None of them sponsor, endorse, or review this site.
The name "WorkBuddy Central" is descriptive of the site's subject, not a claim to anyone's brand.