dsh-Mimir: a research workbench plugin for the whole paper cycle

What dsh-Mimir actually bundles

A research project touches a lot of tools that do not talk to each other: a reference manager, a LaTeX editor, a plotting script, an SSH session to a GPU box, and a folder of meeting slides. Mimir's premise is that the agent driving your work should be able to reach all of them through one plugin. The README describes it as "the research-lifecycle copilot inside DeepSeek Harness", and the nine sidebar views match that claim: Overview, Paper, Library, Experiments, Figures, Meetings, Servers, Ledger and Venues.

The target user is narrow and specific. You need Node.js 22 or newer, the dsh CLI installed globally, and a DEEPSEEK_API_KEY for agent sessions. If you write papers in Word, or your group's compute lives behind a scheduler you cannot SSH into, most of the surface area is dead weight. The plugin is one npm package, dsh-mimir, and it self-activates on install rather than requiring a separate configuration step.

The agent-tool and SSE architecture

Mimir is not a standalone app. It registers agent tools and slash commands inside dsh and renders its UI as an overlay toggled from the sidebar. The README lists the tool names an agent can call: arxiv_search, web_search, wiki_note, figure_save, latex_compile, meeting_deck, venue_search, and a remote-compute quartet of server_list, server_check, server_submit_job and server_list_jobs. The Servers tab and those four tools read the same remembered server and job state, so a job submitted by the agent appears in the UI without a second registration step.

State lives in two places. The wiki persists at ~/.dsh/storages/research_wiki.json, and artifacts land under ./.research by default, configurable through workspaceDir. Every view live-refreshes over SSE as the agent writes, which the README frames as "no manual reload". That is the design bet: the agent is the writer, the views are readers. Eleven bundled skills (literature review, novelty check, experiment planning, citation audit, bilingual de-AI polish, rebuttal) ship with the package and register by default under skills.enabled, so the agent has workflow instructions without you authoring prompts.

Installing dsh-Mimir and opening the workbench

The prerequisite set is explicit: Node.js >= 22, the dsh CLI from npm install -g @deepseek-ai/dsh, and a DEEPSEEK_API_KEY. The install is a plugin add against the web profile, which the README says installs and self-activates.

dsh plugin --profile web add dsh-mimir@latest
dsh web

After dsh web starts, open http://127.0.0.1:3080 and click Mimir in the sidebar footer. The README notes that plugin names resolve from the profile directory, so run the add command as written rather than from your working directory.

Version pinning matters here. The README states that 0.18.x requires dsh >= 0.1.2-alpha.4 because of upstream breaking changes, and that on an older dsh you should pin the previous release with dsh plugin --profile web add dsh-mimir@0.16.0. There is also a pnpm quirk: dsh's plugin store uses pnpm, which holds back freshly published releases, so a recent publish may not resolve through @latest. The documented workaround is to remove and pin an exact version.

dsh plugin --profile web remove dsh-mimir
dsh plugin --profile web add dsh-mimir@0.20.0

Two optional capabilities are worth wiring before you judge the tool. Paper compilation needs a LaTeX engine, and the README calls brew install tectonic the easiest path, with latex.engine accepting an absolute binary path. Web search needs a SearXNG server, and the package ships a setup script for a Docker-free local venv.

bash scripts/setup-web-search.sh

Configuration keys are all optional and go in the profile's cordis.patch.yml, with a fully commented example at examples/mimir-agent/cordis.yml. The defaults worth knowing: workspaceDir is .research, latex.engine is auto, search.command is auto, reviewer.maxRounds is 3, and backup runs every 60 minutes keeping 24 snapshots.

Where dsh-Mimir breaks or is the wrong tool

The dependency on dsh is the first hard boundary. Mimir is a plugin, not an application. Anyone who does not want to run DeepSeek Harness, or who works in an environment where a globally installed CLI and a DEEPSEEK_API_KEY are not available, gets nothing from this repository. The README does point at Mimir-Desktop, a community-maintained Electron edition by @hxhy00 with no dsh install and no backend, but that is a separate project with a separate maintainer, and this article's material does not describe its internals.

Web search is the second boundary. The sxng CLI ships with the package, but it needs a SearXNG server, and the documented path is running scripts/setup-web-search.sh to stand one up locally. That is a real operational cost for a feature that many researchers would expect to work out of the box. The troubleshooting section also lists arXiv failures behind a proxy, resolved by exporting HTTPS_PROXY before starting dsh, which tells you the network assumptions are not invisible.

LaTeX compilation has a first-run penalty the README is honest about: the first tectonic compile downloads packages, and if it times out you raise latex.timeoutMs. There is no documented rollback procedure for a bad compile or a corrupted wiki, and the README does not describe what happens when research_wiki.json is written concurrently. The backup settings (backup.enabled, intervalMinutes, keep) are the only stated recovery mechanism, and they snapshot the wiki, not the artifacts under workspaceDir.

Mimir-Desktop and Overleaf are different answers

The obvious comparison is Overleaf. Overleaf is a hosted LaTeX editor with a collaboration model built around shared projects and a browser. Mimir's Paper view is described as an Overleaf-style studio with edit, compile, PDF preview and a one-click AI fix, but the resemblance stops at the editing surface. Mimir's value is that the same tool call that writes a paragraph can also search arXiv, save a figure, or submit a job to a GPU server. Overleaf has no concept of your SSH fleet. If your work is a shared manuscript with five co-authors who do not run dsh, Overleaf is the correct tool and Mimir is not.

The closer alternative is Mimir-Desktop, the Electron edition named in the README. The difference is architectural rather than feature-level: Desktop removes the dsh install and the backend process, and the README says it carries the same feature set. That makes it the better fit for a researcher who wants the workbench without committing to the dsh agent runtime, at the cost of being maintained by a different person in a different repository. Choosing between them is really choosing whether you want the agent tools and slash commands, which only exist in the plugin.

Maintenance, licence and upgrade cost

The repository is not archived, and the last push was on 2026-09-16. The licence is MIT, which permits commercial and academic use, modification and redistribution provided the copyright notice and permission notice are retained. That is a permissive choice with no copyleft obligation on your own paper or code, though nothing here is legal advice and your institution may have its own rules about bundling third-party npm dependencies.

Upgrade cost is the part to watch. The 0.18.x line required dsh >= 0.1.2-alpha.4 because of upstream breaking changes, which means Mimir's compatibility surface is dsh's release cadence, not its own. The changelog for 0.20.0 shows the kind of churn to expect: compile status backfilled from the on-disk PDF after restarts, bounded exponential backoff for scheduled tasks, a unified remote-job lifecycle where duration-cap kills settle as unknown and capture-cap floods as failed, and experiment metrics gaining min/max directions and zero-baseline charts. Each of those is a behaviour change in state you may already have on disk. Pinning an exact version and reading the changelog before upgrading is the documented way to absorb that.

Editorial conclusion

Adopt dsh-Mimir if you already run DeepSeek Harness, write papers in LaTeX, and want arXiv search, experiment charts and remote GPU jobs in the same place your agent reads and writes. Do not adopt it if you are outside the dsh ecosystem or unwilling to run a local SearXNG instance for web search. Before relying on it, verify that your dsh version satisfies the 0.18.x requirement of dsh >= 0.1.2-alpha.4, that a LaTeX engine resolves through latex.engine, and that ~/.dsh/storages/research_wiki.json is the file your backups actually cover.

Frequently asked questions

Official sources

1. 1692775560/dsh-Mimir-Academic-research on GitHub
2. Issues
3. License: MIT
4. Project website
5. README