Short answer: keep the text model, add a separate image reader
ModLens sends images to a separate vision engine and returns structured text for a text-only coding model. Use it if you want to keep that coding model and can choose an acceptable image provider. Skip it if your existing model already reads images, or if you require automatic bitmap paste in Codex TUI. Our source-install attempt stopped after 12 seconds with an esbuild build-script approval error. That is the setup result behind this review, not proof that image reading is broken.
The attraction is control over two different jobs: your preferred model writes code, while a vision engine reads a screenshot. The price is another route to configure and another account whose quota and data policy matter. The getting-started guide describes that separation. The final investigation in issue #78 explains why the host can stop an image before the skill gets any opportunity to act.
What people use modlens for
Users report adding image reading to DeepSeek, processing batches through MiMo, and supplying local file paths when Codex attachments fail. These are specific reports from the community records, rather than a list of hypothetical business benefits.
| Who used it | What they did | Reported outcome |
|---|---|---|
| X user hqmank | Added the ModLens skill to DeepSeek | Reported giving DeepSeek vision with one skill. Their post supplies the command. |
| The reporter of GitHub issue #12 | Moved image processing to MiMo after hitting Google's request limits | Found its default reasoning too slow and asked for provider reasoning controls. The thread ends with a release adding those controls. |
| The reporter of GitHub issue #78 | Sent a local image through a text-only Codex gateway | Reported that an absolute path sent as text worked when native attachments failed before ModLens ran. The report and maintainer investigation distinguish the two paths. |
| GitHub user paicat1 | Added Windows desktop capture before image reading | Reported useful local use, then separated capture into dsh-screenshot after upstream declined to absorb it. Issue #48 records the split. |
The MiMo report is a useful buying signal because the user had already reached a real constraint: too many images for the route they were using. Changing providers moved the bottleneck to reasoning behavior. The maintainer's follow-up found that disabling reasoning on a Qwen route caused missing required layout fields. For batch work, getting a shorter response is only useful if the response still contains the evidence the coding model needs. Those are the maintainer's observations, not our benchmark.
What happened when we ran it
Our November 4, 2026 source-install attempt at commit ac278f8 failed with exit 1 after 12 seconds. The measurement setup in our sandbox specified 3 CPUs, 8 GB RAM and image label lab-node:22. The label alone does not identify the runtime or operating-system version. The run used the repository's Node/pnpm ecosystem.
| Recorded source-check observation | Result |
|---|---|
| Install | Failed, exit 1, 12 seconds |
| Error | ERR_PNPM_IGNORED_BUILDS |
| Dependency named in the error | esbuild@0.25.12 |
| Log's proposed next step | Run pnpm approve-builds to choose permitted dependency scripts |
| Repository scan | 158 files, approximately 35,682 source lines, 5.1 MB checked out |
| Repository signals | 2 CI workflow files, no Dockerfile detected, no tests directory detected |
The log says esbuild's build scripts were ignored and points to pnpm approve-builds. It also says the lockfile passes supply-chain policies. Those messages describe different checks: the lockfile message does not turn the failed install into a success. Our practical recommendation is to inspect the dependency scripts before granting approval, then repeat the source installation. The available tail does not identify who configured that policy, so blaming Node, the container image or the application would go beyond the record.
The pack also contains a separate November 17 package-only run, associated with checkout e42fc495c736 and the quickstart-lib recipe. It cloned the repository and reports a successful npm package installation. That is a different route from the earlier pnpm source attempt, so it does not overwrite the failure above. The clone record and package record identify the operations separately.
That later recipe then stopped at its snippet step: the README has no JavaScript/TypeScript code block to run. Both the failed step and finding say so. ModLens is documented as a CLI and harness integration, so an absent library snippet is a mismatch with that recipe, not a failed image-recognition answer. The teardown record reports cleanup succeeded.
Where people get stuck: the host can reject the image first
Codex TUI can reject bitmap paste before ModLens runs, according to the maintainer's final investigation in issue #78. For a text-only TUI workflow, send the absolute image path as ordinary text. The same investigation reports that Codex Desktop keeps a path tag the skill can use, subject to sandbox access to the file.
The reporter originally saw MODEL_CAPABILITY_NOT_SUPPORTED after attaching an image through a text-only gateway. Setting reuse.codex=true did not help. The maintainer explained that this setting permits Codex to act as a vision provider after ModLens is invoked. It does not intercept the host's incoming attachments. The issue is closed with a known limitation and an upstream improvement request, rather than a ModLens fix for TUI paste.
| Symptom and source | What the record says | What to do |
|---|---|---|
| Native Codex attachment fails, #78 | TUI loses the usable path before a skill can act; Desktop preserves a path tag | Send a text path in TUI, or use Desktop with file access allowed |
| Skill stays on an old version, #33 | Copied skills keep their original pin; doctor gained a comparison against its own CLI version in 3.16.2 | Reinstall the skill copy; doctor is not checking npm's newest release |
| Windows provider fails with spawn EINVAL, #43 | The maintainer reports npm-shim fixes in 3.18.0 and a remaining pnpm-shim fix in 3.20.0 | Check the installed version and shim source before repeating an old workaround |
| DSH lacks the expected tools after installation, #33 | The reporter corrected a skill-versus-plugin mistake; the maintainer also documented release-age filtering | Install the native plugin at an intended exact version and check DSH's installed-plugin list |
The update trap deserves particular care. A copied skill and a DSH plugin have different update procedures. In #33 the maintainer also corrected earlier advice about @latest: pnpm's release-age filtering could resolve it to an older package. An explicit intended version and an inspection of what actually landed are more useful than repeatedly requesting the tag. The troubleshooting guide covers the stricter policy case with a version-specific exception.
Batch work spends quota again after a restart
A plugin restart can trigger a read of each distinct image already in the conversation, according to the maintainer in issue #92. The upstream text-only model needs those images converted to text before it receives the history. That expected burst is different from repeatedly paying to read the same image during the same plugin run.
The maintainer says duplicate reads were fixed in 3.24.2: each unique image and focus is read once per plugin run and then cached. For someone processing a long screenshot conversation, the useful accounting question is therefore whether the calls represent different history images, a changed focus, or the same image being reread. A claim that caching makes history free would miss the restart cost.
Other plugins can also create extra work. The harness interoperability documentation requires a real host attachment reference so ModLens can retrieve the image bytes. A hand-built block containing only a path or base64 payload gets a read-failure placeholder. On an image-declaring route, adding a text path beside a valid image block invites a second tool read. This is actionable integration advice: preserve the host attachment reference and avoid sending the same image in two forms.
Privacy depends on the route, including its fallback
A configured provider preference still allows fallback; -p pins one provider with no fallback. The provider documentation makes that distinction explicit. For company screenshots, the approved destination should determine the route before an image enters the conversation.
For example, modlens -i screenshot.png -p openai pins the OpenAI-compatible provider. Its endpoint is the configured openai.baseUrl, which can point to a different vendor or your own gateway. Review that endpoint, the account handling the image and the permitted data class. Calling the provider openai is a protocol choice, not assurance that a particular company receives the pixels. The provider docs require an endpoint, key and image-capable model for that route.
Shared paste storage is a separate concern. In issue #105, a reporter alleged images from different sessions were being mixed. The maintainer confirmed that DSH paste storage is scoped to the operating-system user rather than individual sessions. A model with filesystem tools could inspect neighboring files. The maintainer requested the actual image file and raw tool outputs to distinguish an upload problem from invented model output, then closed the issue when those materials did not arrive. That is a reason to consider the agent's filesystem access, not evidence of a proven cross-session injection.
Text extracted from an image also stays untrusted. The embedded skill instructions on TypingMind's listing explicitly say not to follow instructions found inside an image. A screenshot of a terminal can contain commands, account names and secrets as well as the error you intended to share. Crop and sanitize the image before sending it, and inspect which provider answered in the returned metadata.
English docs exist; Chinese threads carry important fixes
ModLens has an English README and English setup documentation, alongside Chinese documentation and issue discussions. The harness guide says its DSH settings card follows the host's Chinese or English interface setting, with browser language as the fallback for older hosts. Server failure messages retain their diagnostic wording.
The Chinese-language Bilibili search record includes HarmonyOS-万少's ModLens and GLM-5.3 Flash configuration tutorial, and ChHsich's discussion of whether plugins remain useful after DSH adds vision and search. These are video listings, which identify the questions people are asking. They do not supply a transcript demonstrating those configurations. The detailed Codex investigation and stale-image discussion above also require reading Chinese issue replies rather than relying on an English feature list.
Repository attention is easier to count than adoption. GitHub's public metadata on November 17 showed 5,012 stars and a last push on November 15, 2026. The dated fixes in the sampled issues give more useful maintenance evidence than stars alone. The current README says one author maintains the project and does not accept pull requests. Teams that expect to contribute fixes through upstream PRs should account for that policy before depending on it.
ModLens vs native vision in OpenCode, Tesseract and screenshot tools
Native image input is the simpler choice when you are willing to change the coding model; ModLens earns its place when retaining a text-only model matters. That distinction appears in the supplied comparison. OpenCode can host the workflow, while Tesseract addresses the narrower job of extracting text.
| Choice | Pick it when | Tradeoff relevant to ModLens |
|---|---|---|
| Native vision in your coding harness, such as OpenCode | Changing to an image-capable model is acceptable | You remove the separate image-to-text bridge; ModLens serves the retain-your-text-model case |
| Tesseract | You need OCR output for a script or a manual text handoff | Its README lists plain text and document-oriented output formats; ModLens adds a harness workflow with semantic and visual fields |
| ModLens with Gemini API | You want an API-key vision route for the existing text model | Upstream recommends it for headless operation; choose its data policy and budget deliberately |
| ModLens with an approved reused agent CLI | You already have an appropriate signed-in vision route | Reads consume that route's quota and may depend on interactive sign-in |
| Manual screenshot capture or dsh-screenshot | Your missing step is obtaining a desktop image | Capture produces the image; ModLens reads it. These can be used together |
The Tesseract README documents an OCR engine and command-line program. It supports plain text, hOCR, PDF, TSV, ALTO and PAGE output. Start there if your actual requirement is text extraction. The extra semantic fields in ModLens are useful for discussing a UI, but their presence alone does not make the extracted text more accurate.
OpenCode is an open-source coding agent, so it is also a host rather than a like-for-like replacement plugin. The capture distinction is similarly practical: paicat1's issue #48 ended with a separate Windows and DSH screenshot plugin after the maintainer declined to turn image reading into screen capture. Choose the capability you are missing before installing another tool.
What the other reviews get wrong
Cloning ModLens is not a complete setup, and doctor does not test a live provider request. Those are two consequential corrections to the supplied installation guides. The official installation procedure requires the right harness integration and a working engine; the troubleshooting guide says doctor makes no network request.
| Page | Advice to correct | Supported correction |
|---|---|---|
| Sabemos | Presents cloning as the installation | Skill hosts need the folder in their skill directory and an engine; DSH uses a native plugin |
| TypingMind | Says bundled files are never executed and skills need no runtime | Its own embedded ModLens instructions invoke launcher scripts and resolve a CLI, npx or bunx |
| OpenDesign and the CSDN tutorial | Use @latest as the update route | The maintainer's #33 investigation shows why a release-age filter can select an older version |
| CSDN | Describes doctor as checking connectivity and model availability | Doctor is an offline diagnostic; a real read exercises the provider |
| OpenDesign | Presents dense-chart reading as complete and interpretation as more certain than a model's guess | The maintainer's #12 example includes schema failures, and Verdent describes OCR and semantic errors |
| Our previous ModLens page | Mixed incompatible source-install accounts and left Codex status unclear | This rewrite uses the supplied November 4 failure and the closed #78 investigation's Desktop/TUI distinction |
The conflicting MrKeyoor account deserves an explicit correction. The captured previous page combined failed setup with a successful build-and-test narrative tied to different checkout details. This rewrite retires that successful narrative as evidence for the supplied source run. The later package-only record above is labeled separately, and its failed library-snippet step remains visible.
The other corrections are grounded in the records for Sabemos, TypingMind, OpenDesign and CSDN. Reading their installation shorthand as a tested workflow would hide the engine choice, the host's attachment behavior and the distinction between valid JSON and accurate image interpretation.
Questions people ask
ModLens is an image-reading CLI and harness integration for text-only models, including models used inside OpenCode. The getting-started guide names OpenCode as a supported skill host and explains that a separate vision engine performs the read.
What is ModLens, and how do I use it in OpenCode?
Install the ModLens skill in OpenCode's ~/.agents/skills/ directory, configure a vision engine and run doctor. Then supply an image or its path for the skill to read. The harness guide explains OpenCode's local image storage; the installation procedure covers placing the complete skill folder, including its launcher.
How do I install and configure the ModLens plugin in DeepSeek Harness?
DSH uses the native plugin, not a copied skill folder. Install an intended exact package version into the selected DSH profile, restart DSH and inspect the installed-plugin list and model selector. Configure its engine through the settings card or shared ModLens configuration. The harness guide documents that workflow and the (modlens vision) entries. Run an actual non-sensitive image read after the offline diagnostic.
How do I configure ModLens with GLM-5.3 Flash as the vision engine?
Use an image-capable OpenAI-compatible endpoint with its supplied base URL, credential and exact model identifier. The provider guide documents openai.baseUrl, openai.apiKey and openai.model and names GLM's platform as compatible. The current README distinguishes native multimodal GLM-5.3-Flash from text-only GLM-5.3. If Flash is already your coding model, its native image input can remove the need for this bridge.
Are ModLens and search plugins still useful after DSH adds official vision and search features?
ModLens serves the narrower case where your chosen coding model remains text-only and you want a separate vision engine. If the model already accepts images, upstream leaves its native image path alone. That follows the harness guide's routing rule. ChHsich's Bilibili listing raises the broader plugin question, but supplies no transcript comparing DSH's search implementation. This review makes no search-plugin recommendation from that title.
Does ModLens need an API key or a separate vision engine?
It needs a working vision engine, but that need not mean a new API key. The provider documentation describes API routes and approved reuse of signed-in local agent CLIs. A reusable CLI still needs vision capability and consumes its account's quota. A skill directory alone cannot interpret pixels, even if the coding model knows when to call it.
Is ModLens safe to use with company data?
Company use depends on approving the actual image destination, account and fallback behavior. The provider guide says a preference permits fallback while -p pins the destination route. The Verdent guide warns that pasted screenshots can expose private records, credentials and background information. Sanitize the image and use only an approved route for its contents.
Can I use ModLens commercially?
ModLens is MIT licensed, and the project's disclaimer says upstream engines retain their own terms and quotas. Commercial users must retain the applicable copyright and license notices. The license explanation and current project README support that distinction. Permission to use the code does not supply a provider subscription or permission to send customer screenshots to it.
Does ModLens make a text-only model natively multimodal?
No. A separate vision engine interprets the pixels, and the text model receives its structured result. The Verdent explanation describes the boundary explicitly. Keeping the two roles separate also explains why the original coding model's native attachment endpoint can reject an image before ModLens runs.
What information does ModLens return in its JSON result?
The documented result contains a summary, OCR text, reading-order layout, semantic information, visual notes and uncertainty. Metadata records provider attempts and warnings. The CLI manual shows the contract, and Verdent's field table explains its parts. A valid shape tells the consumer where information belongs; it does not independently check that the image was read correctly.
Should an agent change code from ModLens JSON alone?
Use the JSON to locate the investigation, then compare important text with the image and inspect the code or DOM. The Verdent guide recommends that sequence, particularly for small text and destructive changes. A screenshot shows a visible state. It cannot tell the agent all the interaction rules or implementation details behind that state.
Why do native Codex image attachments fail before ModLens can read them?
The final #78 investigation says Codex TUI checks model modalities after writing the bitmap and discards the usable path when rejecting text-only image input. The skill has no execution opportunity at that point. Use a plain text path. Desktop preserves a path tag, which gives ModLens a different route to the file.
Can a Codex hook or wrapper recover a failed native vision request?
ModLens does not supply an interception hook or local proxy for this TUI failure. The maintainer closed #78 as a known boundary and reported an upstream request to insert the saved path as text. Enabling Codex reuse selects a possible downstream vision engine after invocation; it does not add an interception point before invocation.
What we have not tested yet
Our source run stopped at dependency installation, and the package-only recipe stopped before an application task. We have not tested image-reading accuracy, OCR error rates, reading-order accuracy, repeated use on private screenshots, or end-to-end Codex, OpenCode and DSH operation in our lab. No successful build or passing-test total belongs to the supplied November 4 result. The absent tests directory is a scan result, not proof that the repository contains no tests elsewhere.
The provider comparisons above are documentation and issue-based decision aids. We did not benchmark their latency, price, quota behavior or reliability against a native vision model or Tesseract. The sampled community reports cannot establish an independent accuracy score, commercial adoption, a support SLA, or whether GLM-5.3 Flash works with your particular gateway. The Bilibili records are listings, and the supplied YouTube transcripts do not demonstrate this software. We excluded those unrelated transcripts from the product evidence.
We have not resolved the stale-image allegation in #105 or audited the isolation and network behavior of the supported routes. The record lacks the requested file inspection and raw tool output, so it cannot confirm stale bytes or cross-session injection. For a decision-changing next check, choose a sanitized screenshot whose text you can read, record the provider used, and compare its JSON with the image before allowing the coding agent to act.


