DeepSeek Harness plugins worth installing
Five DeepSeek Harness plugins that change how a rented server behaves: budget caps, tool permission rules, injection scanning, durable memory, LAN access.
Which DeepSeek Harness plugins are worth installing
DeepSeek Harness plugins are third-party code that runs inside your agent, with your agent's privileges, on a machine you pay for. The community list carries more than twenty categories and over a thousand entries. On a rented virtual private server (VPS) you need five: dsh-budget for spend caps, dsh-permission-rules for tool gating, dsh-defend for injection and secret scanning, dsh-memory for facts that outlive a session, and dsh-web-lan-access only after you have decided how you will authenticate.
The harness is dsh, DeepSeek's open-source agent harness, built so that everything is a plugin. Its own README calls it a developer preview and warns THERE WILL BE COMPATIBILITY-BREAKING CHANGES. That one fact shapes every choice below. Pin what you install, expect an upgrade to break it, and keep the set small enough that you can actually read all of it. If the harness is not running yet, start with a DeepSeek Harness install on a VPS and come back here. If the parts these plugins hook into, the agent loop, its tools, its memory, are still fuzzy, work through the fundamentals first, because every choice below is easier to judge once you know what each layer is doing.
How dsh installs a plugin, and where the config lands
dsh assembles itself from Cordis plugins, so a running instance is a tree of plugins rather than one program. A profile is a named composition of those plugins, stored in the harness home. The harness ships five profile templates, web, headless, sdk, sdk-minimal and acp; the two that matter on a server are web and headless. $DSH_HOME defaults to ~/.dsh, and a profile lives at $DSH_HOME/profiles/<name>/. It declares itself in its own package.json under a dsh field, where dsh.profile lists the profile's bundles, and the same directory holds any out-of-tree plugins it installs and the user's own cordis.patch.yml.
dsh plugin --profile web list
dsh plugin --profile web add dsh-budget
dsh plugin --profile web remove dsh-budgetFour source forms work: a bare npm package name, a scoped npm name such as @scope/plugin, a GitHub reference such as github:PerryLink/dsh-budget#main, and a local path with link: or ./. Reach for the GitHub form, because you can replace #main with a commit and get the same code back next month.
dsh plugin --profile web add "github:PerryLink/dsh-budget#461d478"Layers apply in a fixed order: each bundle in the profile's listed order, then the profile's cordis.patch.yml, then the home-level cordis.patch.yml, then any --patch overlay. Order matters because a later layer can reconfigure or drop what an earlier one set up. So when a plugin looks installed but does nothing, check that its bundle is selected in the profile's package.json, then check that no later patch layer disables it; the insert block normally comes from the bundle's own patch file.
As of 17 August 2026 the npm package @deepseek-ai/dsh was at 0.1.0-rc.7, while the plugins below were built for the 0.1.0-rc line. At the commits pinned here, dsh-permission-rules and dsh-defend declare compatibility with 0.1.0-rc.5 to 0.1.0-rc.6, dsh-budget declares 0.1.0-rc.6 alone, dsh-memory sets only peer ranges starting at ^0.1.0-rc.6, and dsh-web-lan-access declares no dsh version at all. On 16 October 2026 the latest tag on npm was 0.2.0-rc.2, and on 20 October 2026 it was 0.2.1-rc.2, which it still was on 27 October 2026, when the next tag had moved on to 0.2.1-rc.4. Checked 10 November 2026, npm's latest and next tags both point at 0.2.2-rc.1. So those pinned commits now trail the harness by a whole minor line. Most of the plugins have since caught up on their main branches. Checked on 16 October 2026, dsh-budget 0.4.17, dsh-permission-rules 0.7.15 and dsh-defend 0.3.21 each accept dsh from 0.1.2-rc.1 up to, but not including, 0.3.0, and each lists 0.2.1-alpha.1 among the harness releases it was tested on. dsh-web-lan-access 1.4.0, released 17 October 2026 and checked 18 October 2026, declares >=0.2.1-rc.1, with tested releases listed up to 0.2.1-rc.2, which npm carries on both its latest and next tags as of 20 October 2026; its README pairs harness 0.2.0-rc.x and earlier with plugin 1.3.x. Checked 10 November 2026, the same 1.4.0 package file also lists 0.2.2-rc.1 among its compatible releases. dsh-memory on main, checked 27 October 2026, still declares ^0.1.0-rc.6 for both @deepseek-ai/dsh-tools and @deepseek-ai/dsh-system-prompt. That range still covers the 0.1.0-rc builds the pins were chosen for, from 0.1.0-rc.6 onward, but not the 0.2 line. None of the newer ranges on the other four plugins covers those 0.1.0-rc builds, so moving one of those pins to current code means moving the harness with it. That gap is normal in this ecosystem, and it is the usual reason a plugin stops loading: the harness moves faster than the plugins around it. Checked 29 October 2026, the 0.2.1-rc.2 CLI enforces declared harness compatibility during installation and startup, so the older pins are refused on an incompatible harness unless you explicitly approve an exact-version exemption. Upgrade the harness on purpose, then test one plugin at a time.
Read a plugin before you trust it
A dsh plugin is not sandboxed from the harness. It loads into the same Cordis tree, in the same process, as the same operating system user, with access to the same model credentials and the same working tree. Installing one is closer to running someone else's script with sudo than to adding a browser extension. That is the same trust question Claude Code plugins raise, and the answer is the same: read the code, or do not install it.
Four things to check, in this order:
- Which extension points it taps.
tools/pre-executemeans it sees every tool call and can block it.agent/pre-stepmeans it sees your messages.webServer.tapIndexmeans it rewrites the page served to your browser. A plugin tapping none of those can do very little; a plugin tapping all of them is your security boundary. - Whether it talks to the network. Search the source for
fetch,http, and any hard-coded hostname. A cost meter that phones home is shipping your usage pattern somewhere. - Whether it reads credentials. Anything touching
credentials.*or a provider key needs a stated reason in the README. - The license and the last commit date. An unlicensed plugin, untouched for months, in an ecosystem that changes weekly, is a liability.
Then install by commit rather than by branch, confirm the bundle selection in the profile's package.json, and read the insert block in the bundle's own patch file. That block names the plugin id and the config it registered, which is the shortest honest description of what you just added. Keep provider keys out of a plugin's reach where the harness lets you, along the lines of keeping secrets out of AI agents.
dsh-budget: how do I stop an agent spending all night?
An agent on a VPS runs unattended. That is the point of putting it there, and it is also the risk. dsh-budget meters tokens and estimated cost per model, per session and per day, and enforces caps against those totals.
Its current README, checked 10 November 2026, marks feature development frozen since 5 October 2026, with fixes reserved for genuine breakage. Existing installs remain supported, but the maintainer recommends dsh-cost-meter for new work.
dsh plugin --profile web add "github:PerryLink/dsh-budget#461d478"Its defaults are generous: 10 USD per session, 50 per day, and 500 per month. Those suit a funded team. On a personal server they are high enough that a runaway loop finishes before the cap ever fires, so lower them on the first day.
The data behind this chart
[
{
"label": "Per session",
"plugin_default_usd": 10,
"suggested_start_usd": 2
},
{
"label": "Per day",
"plugin_default_usd": 50,
"suggested_start_usd": 5
},
{
"label": "Per month",
"plugin_default_usd": 500,
"suggested_start_usd": 40
}
]The suggested column is a starting point for one operator on one box, not a published figure. Raise it when a real month of usage tells you to. Lower caps on their own only change when you get notified: the pinned configuration ships overLimit: alert, which it describes as notify only. A monthly cap of 40 USD with a session cap of 2, together with overLimit: block as in the config below, fails loudly and early, which is what you want while you are still learning the tool's appetite.
- id: budget
config:
budgets:
session: 2
daily: 5
monthly: 40
warnRatio: 0.8
overLimit: blockThe setting that changes how the machine behaves is overLimit. Its default is alert, which prints a warning and keeps spending, so out of the box this plugin is a dashboard. Set it to block and the harness refuses further model calls once a cap is reached, which means an overnight job stops instead of billing until morning. degrade is the third mode: it blocks the call with guidance naming a cheaper model from the degradation map, which is right when a job must finish but does not need your best model. warnRatio defaults to 0.8, so you hear about it at 80 percent of the cap.
Two honest limits. Cost is computed from a built-in price table merged with the prices you supply, and the maintainer warns that built-in prices drift. prices is empty by default. At the pinned commit 461d478, defaultPrice, the fallback for a model in neither table, is 1.0 USD per million input tokens and 3.0 per million output; on current main, checked 10 October 2026, it defaults to zero and the model surfaces as unpriced. Fill in the real numbers for your models or the dashboard is a guess dressed as a measurement. Second, at 461d478 the plugin aggregates in the running process from the session event stream, so the totals reset when the harness restarts, and a crash loop, or a supervisor that restarts dsh, resets a daily cap. Current main persists the day and month buckets by default (persistence.enabled: true, snapshotted every 10 seconds) and restores them at startup, so a restart no longer clears a daily or monthly cap there, provided the harness supplies the storage domain; without it, aggregation stays process-local. main also targets the 0.2 harness line, so moving the pin means moving the harness too. Either way, treat dsh-budget as a guardrail against your own jobs and keep a spend limit on the provider account as the actual ceiling, which is the wider argument in controlling AI agent cost on a VPS.
Day to day you use /budget for the overview, /budget models for the per-model breakdown, and /budget unblock <scope> to lift a block once you have decided to keep going.
dsh-permission-rules: which tool calls should never run?
dsh-permission-rules puts declarative rules on the tools/pre-execute waterfall, so a rule is evaluated before a tool runs. Three actions exist. allow passes the call through, deny blocks it and returns a reason the model can read, and ask routes it to the official approval seam.
dsh plugin --profile web add "github:PerryLink/dsh-permission-rules#b30b4fb"Rules live in .dsh/rules.yaml relative to the session working directory, with a global fallbackPath and optional searchUp to walk up toward the filesystem root. Matching covers tool-name globs, parameter key and value globs, workspace-relative path globs, agent selectors such as main or subagent, and network targets.
rules:
- match: { tools: [bash], params: { command: "rm -rf*" } }
action: deny
reason: "No recursive deletes"
- match: { tools: [edit, write], paths: ["**/.env*", "**/secrets/**"] }
action: ask
reason: "Secret files need confirmation"
- match: { tools: ["mcp__*"] }
action: ask
reason: "MCP tools need confirmation"Evaluation is first match wins, so a broad allow near the top quietly cancels every narrower rule under it. Write the denials first and the permissive rules last. The tool-name glob covers mcp__*, which is how you gate tools that came from a model context protocol (MCP) server rather than from the harness itself, and that matters as soon as you are running MCP servers on a VPS.
One behaviour to plan for: ask needs an answerer. On a headless profile nobody may be watching, and an unanswered ask does not wait for someone to appear: with no answerer composed, the harness approval chain fails closed and the call is refused. On an unattended box an ask rule therefore behaves like a deny you did not write on purpose. Use deny for anything you would never approve, keep ask for the profile you sit in front of, and if you want approvals to work unattended you need a real answering path, which is the subject of gating AI agent actions with approvals.
dsh-defend: what about prompt injection and leaked secrets?
dsh-defend scans at three points: inbound messages on agent/pre-step, tool arguments on tools/pre-execute including a destructive-delete guard, and tool results on tools/post-execute. That last one is the interesting one, because it is where content fetched from a web page gets a look before the model acts on it.
Its current README, checked 10 November 2026, marks feature development frozen since 5 October 2026, with fixes reserved for genuine breakage.
dsh plugin --profile web add "github:PerryLink/dsh-defend#7ba3427"Defaults are cautious rather than strict: detection.injectionAction, detection.jailbreakAction and detection.secretAction are all ask, while detection.secretBlockCritical is true, so a critical secret is blocked whatever the other settings say. On an unattended box change the actions you care about to block, because ask with nobody to ask is not a decision.
The audit design deserves credit. defend/detection events record rule id, family, category, severity, decision and scan facts, and never the matched text; secrets are recorded by type only. So enabling the audit log does not create a second copy of the credential you were trying to protect.
Be clear about what this buys you. Detection is rule-based, and the README says plainly that novel phrasing and multi-step attacks can get past it. It lowers the rate of obvious attacks. It does not make an agent safe to point at untrusted content, so keep the permission rules underneath it.
dsh-memory: what does the agent remember tomorrow?
At least three different plugins are called dsh-memory, which is worth knowing before you type an install command. Install by explicit source so you get the one you read. Two are compared below. The third, documented as @dsh-memory/bundle, is a different design: checked 20 October 2026, it keeps markdown files under ~/.dsh/memory/<project>/ and ~/.dsh/memory/_user/, has a model distill them from session logs, injects a token-capped index of them, and offers only memory_read and memory_expand as tools. Nothing below describes it.
The one I would put on a small server is the SQLite build. It registers as memory, keeps one file at $DSH_HOME/memory/memory.db shared across profiles, and exposes memory_write, memory_search and memory_forget. Search is keyword-based over the stored text and tags. There is no embedding service, no API key, and no extra process.
dsh plugin --profile web add "github:ben7am1n/dsh-memory#def7c6a"The configuration you will touch is small: path for the database file, promptRecentCount (default 10) for how many unpinned memories get injected, and promptMaxChars (default 2000) for the rendering budget. It leans on node:sqlite, which Node 22 still marks experimental and Node 24 only promoted to release candidate in 24.15.0 (checked 10 November 2026), so a Node upgrade is a thing to test rather than assume.
Here is what memory actually changes about the machine: injected memories ride in the system prompt on every turn. promptMaxChars is a budget, not a fixed amount, and not quite a hard ceiling either: on main, checked 27 October 2026, the header and memory lines are counted against it, but the notice saying how many memories did not fit is appended after the check, so a 2000-character setting can render a few dozen characters more. Once your stored memories fill it, a 2000-character budget is a few hundred extra input tokens on every single request, forever. That is a real line on your bill, and it is why dsh-budget belongs on the box before dsh-memory does. Keep promptMaxChars tight and prune with memory_forget instead of letting the file grow.
The alternative build stores memories in a YAML file and does embedding search with automatic prompt injection, installed with dsh plugin --profile web add github:towzai/dsh-memory. Checked 10 November 2026, its default embedding search needs a local ollama instance and an embedding model, qwen3-embedding:0.6b by default, overridable with DSH_MEMORY_EMBED_MODEL; without ollama it falls back to keyword-only search, and an embedProvider setting of nous sends embedding requests to an OpenAI-compatible cloud endpoint instead. Semantic recall is better than keyword recall. With the local default it also means a second service and a set of model weights resident on the same server, which on a small plan is memory taken from the work you rented the box to do. Pick it when you have RAM to spare. The general tradeoff between recall quality and resident cost is covered in local memory for agents.
dsh-web-lan-access: should the web UI listen beyond loopback?
npx @deepseek-ai/dsh web serves the interface on 127.0.0.1:3080. Browsers only expose crypto.randomUUID() in a secure context, so on harness builds up to the 0.2.0 line, loading that same page over plain HTTP from another machine fails. dsh-web-lan-access 1.3.x fixed that by tapping webServer.tapIndex to inject a small polyfill, and switched the server binding to 0.0.0.0. Version 1.4.0, released 17 October 2026 for dsh 0.2.1, removed the polyfill: that harness generates UUIDs with crypto.getRandomValues, so a plain-HTTP LAN page renders without a patch. The plugin still taps webServer.tapIndex, now to inject a bootstrap script that lets a remote page use Host-backed settings, and it still binds 0.0.0.0. On 0.2.1 the harness itself refuses a wildcard bind, warning that binding every interface would expose remote code execution to the network, and the plugin replaces the shipped webserver to get past that check. Checked 18 October 2026.
Read its own warning before you install it. At the commit cited here, e27e909, the README said that binding 0.0.0.0 makes the agent reachable without authentication by anyone on the same local area network (LAN), and on a server with a public IP address the whole internet, with a short list of sensitive methods (settings.*, credentials.*, llm.discoverModels) pinned to loopback and returning 403 from remote origins. That has changed. Current harness builds, checked 10 October 2026, authenticate the browser: the launch token in the printed URL is exchanged for a signed session cookie, which every Host API route and stream requires. The plugin's current README calls trustedHosts an Origin/Host fence, not identity, says binding 0.0.0.0 exposes the authentication surface on every reachable interface, and now enables Host-backed settings for an authenticated remote page, so on current builds a signed-in remote browser can reach Host settings too. The risk has moved rather than gone: anyone who can reach the port can probe the login, the printed URL is itself a credential, and over plain HTTP that token crosses the network in cleartext.
Most of the time you do not need this plugin at all. Forward the port over SSH instead.
ssh -N -L 3080:127.0.0.1:3080 you@your-serverThen open the URL dsh printed when it started, ?token= part included, in your local browser at http://127.0.0.1:3080. Checked 27 October 2026, that first visit trades the token for a signed session cookie that lasts 30 days by default, and after it the bare http://127.0.0.1:3080 works. A browser without a valid cookie gets a 401 on the bare address, and the token changes every time the harness starts, so a fresh browser needs the URL from the latest start. The harness keeps listening only on loopback, so nothing is exposed, and because browsers treat 127.0.0.1 as a secure origin, crypto.randomUUID() is available and no polyfill is needed. One command, no plugin, no new attack surface.
Install the plugin only when a forward will not do, for example when a phone on the same network has to reach the UI. If a single private interface is enough, the plugin's current README recommends dsh web --host <ip> instead; that binds the chosen address without a plugin. Checked 30 October 2026, the harness also supports --tls-cert and --tls-key together to serve HTTPS directly with your own certificate and key. If you do install it, keep a firewall rule that allows access only from your private network, and list any hostnames you will use under trustedHosts. On plugin 1.4.0 that list sits on the connection entry, and a literal list replaces the interface addresses the plugin derives rather than adding to them, so keep those IPs in it. For genuine multi-user access there is dsh-passwords, which adds a login, subuser permissions, per-subuser token and time caps, automatic TLS (transport layer security) certificates through Let's Encrypt on host installs, an audit log, and encryption of sensitive fields in its SQLite store. Judge it as a platform rather than a plugin. Its recommended path, checked 17 October 2026, is a Docker container published on loopback only (-p 127.0.0.1:3088:3088), which serves plain HTTP and expects a reverse proxy in front of it for HTTPS. Automatic HTTPS belongs to the host install and needs public ports 80 and 443; port 80 is not needed if you supply your own certificate through MCP_GATEWAY_TLS_CERT and MCP_GATEWAY_TLS_KEY. The host install's quick path pipes a shell script from the network straight into sudo bash. If you install on the host, prefer npm install -g dsh-passwords followed by sudo dsh-passwords install, so the code sits on disk where you can read it before it runs.
How to remove a plugin cleanly
Uninstalling is two steps, and people skip the second one.
dsh plugin --profile web remove dsh-budget
dsh plugin --profile web listlist should no longer show it. Then check that the profile's package.json no longer selects its bundle, and delete any manual insert block for it from $DSH_HOME/profiles/web/cordis.patch.yml; bundle-provided rows disappear when the bundle is removed. Restart the harness so the tree is rebuilt; with HMR enabled, configuration changes can apply before a restart. Finally, remember that data outlives code. $DSH_HOME/memory/memory.db and .dsh/rules.yaml survive an uninstall. Delete them yourself if removing the data was the point.
What I read, and when
Every reference here is a commit, not a branch, because main will be different code by the time you read this. I read all of it on 17 August 2026, and the harness itself was at 0.1.0-rc.7 on npm that day. I rechecked the defaults and behaviour described above against the primary sources on 10 October 2026, when the latest tag on npm was 0.2.0-rc.2, and rechecked the plugins' declared harness compatibility on 16 October 2026, when it still was. On 17 October 2026 I rechecked the profile layout against the harness architecture doc, both dsh-memory READMEs and the dsh-passwords README; latest was still 0.2.0-rc.2 that day. On 18 October 2026 I rechecked both dsh-memory builds and the current dsh-web-lan-access package and README; latest was still 0.2.0-rc.2, with 0.2.1-rc.2 on the next tag. On 20 October 2026 I rechecked the SQLite dsh-memory README at def7c6a and its main peer ranges, the YAML build's README, the Node node:sqlite docs and npm's dist-tags; latest and next both pointed at 0.2.1-rc.2. On 27 October 2026 I rechecked the SQLite dsh-memory package file, source and README on main, the dsh-budget configuration and source at 461d478, the YAML build's README, the harness note on browser token authentication and npm's dist-tags; latest was still 0.2.1-rc.2 and next was 0.2.1-rc.4. On 29 October 2026 I rechecked all three memory plugins' READMEs, the SQLite build's source and peer ranges, the pinned budget configuration, the current budget configuration and maintenance notice, the pinned permission and defense READMEs, the current plugin package files, the LAN plugin's current and pinned READMEs, the passwords README, the Node SQLite docs, and the harness installation and browser-authentication docs. The SQLite build still uses FTS5 keyword search without an embedding service; the separate Towzai build still documents Ollama embeddings, keyword fallback and the nous provider. npm's latest package endpoint still returned 0.2.1-rc.2.
On 30 October 2026 I rechecked those memory READMEs, the SQLite source and peer ranges, the pinned and current budget configuration and source, the pinned permission and defense READMEs, the current budget and defense maintenance notices, the current plugin package files, both LAN READMEs, the passwords README, the Node SQLite docs, and the harness architecture, plugin-management, browser-authentication and Web deployment docs. The SQLite build has no embedding service or embedProvider setting; the Ollama and nous settings belong to Towzai's separate YAML build. npm's latest package endpoint returned 0.2.1-rc.5.
On 2 November 2026 I rechecked all three memory READMEs, the SQLite source and peer ranges, the pinned and current budget configuration and source, the pinned permission and defense READMEs, the current budget and defense maintenance notices, the current plugin package files, both LAN READMEs, the passwords README, the Node 22 and 24 SQLite docs, and the harness architecture, CLI, browser-authentication and Web deployment docs. The SQLite build still provides keyword search without an embedding service, model weights or an embedProvider setting. Towzai's separate YAML build still documents Ollama embeddings with qwen3-embedding:0.6b, DSH_MEMORY_EMBED_MODEL, keyword fallback and the nous provider. npm's latest package endpoint returned 0.2.1-rc.5.
On 3 November 2026 I rechecked the SQLite dsh-memory README and plugin entry point on main, Towzai's YAML build README, the current budget and defense maintenance notices, the current LAN plugin package file, the Node 22 and 24 SQLite docs, and npm's dist-tags. The SQLite build still describes memory_search as keyword search over memory text and tags, with no embedding service, no API key and no sidecar process. Towzai's separate YAML build still documents hybrid search with local Ollama embeddings, qwen3-embedding:0.6b, DSH_MEMORY_EMBED_MODEL, keyword fallback and the nous provider. npm's latest and next tags both returned 0.2.1-rc.5.
On 10 November 2026 I rechecked the SQLite dsh-memory README and plugin entry point on main, Towzai's YAML build README, the current budget and defense maintenance notices, the current LAN plugin package file and README, the Node 22 and 24 SQLite docs, and npm's dist-tags. The SQLite build still stores memories in one local SQLite file with an FTS5 index and describes memory_search as keyword search over memory text and tags, with no embedding service, no API key and no sidecar process. Towzai's separate YAML build still documents hybrid search with local Ollama embeddings, qwen3-embedding:0.6b, DSH_MEMORY_EMBED_MODEL, keyword fallback and the nous provider. npm's latest and next tags both returned 0.2.2-rc.1.
Exact commits behind this shortlist
- The community plugin list at
f2918fb, 17 August 2026. Linked once on purpose. It is a directory, and a directory is not a recommendation. - dsh-budget at
461d478, 17 August 2026. Apache 2.0. Declares dsh0.1.0-rc.6, Node 22.19 or 24 and up. - dsh-permission-rules at
b30b4fb, 17 August 2026. Apache 2.0. Declares dsh0.1.0-rc.5to0.1.0-rc.6. - dsh-defend at
7ba3427, 17 August 2026. Apache 2.0. Declares dsh0.1.0-rc.5to0.1.0-rc.6. - dsh-memory at
def7c6a, 13 August 2026. MIT. SQLite build. No dsh compatibility field, only peer ranges from^0.1.0-rc.6. - dsh-web-lan-access at
e27e909, 16 August 2026. MIT. No dsh compatibility field.
Recheck those pins before you copy a command. In a developer-preview ecosystem, a version number with a date on it is the only kind that means anything.
FAQ
Which DeepSeek Harness plugins should I install first on a VPS?
Install dsh-budget and dsh-permission-rules before anything else. A budget with overLimit: block stops an unattended run from spending all night, and a .dsh/rules.yaml file stops a tool call you would never have approved. Add dsh-defend once the agent reads anything from the public web, and dsh-memory when you notice yourself pasting the same context into every session. Skip the themes and status chips: they add code running with your agent's privileges without changing what the machine does.
Are dsh plugins sandboxed from the harness?
No. A plugin loads into the same Cordis tree as the harness, in the same process, as the same operating system user, with the same model credentials and the same working tree. A plugin tapping tools/pre-execute sees and can block every tool call; one tapping agent/pre-step sees your messages. So read the source, check the license and the last commit date, and install by commit rather than by branch so the code cannot change under you.
Will dsh-budget actually stop the agent, or only warn me?
That depends on overLimit. The default is alert, which warns at warnRatio and keeps spending. block refuses further model calls once a cap is reached, and /budget unblock <scope> lifts it when you choose to continue. degrade blocks with guidance naming a cheaper model from the degradation map. One limit matters: at the pinned commit 461d478, totals are aggregated in the running process, so restarting the harness resets them and a restart loop will defeat a daily cap. Current main persists day and month buckets across restarts by default. Either way, keep a spend limit on the provider account as the real ceiling.
How do I remove a dsh plugin completely?
Run dsh plugin --profile web remove <package-name>, confirm with dsh plugin --profile web list, then check that the profile's package.json no longer selects its bundle and remove any manual insert block for it from $DSH_HOME/profiles/web/cordis.patch.yml. Restart the harness so the plugin tree is rebuilt. Data written by the plugin stays on disk: $DSH_HOME/memory/memory.db and .dsh/rules.yaml both survive the uninstall until you delete them yourself.
Is it safe to expose the dsh web UI over the network?
Not on its own. dsh web listens on 127.0.0.1:3080, and dsh-web-lan-access changes that binding to 0.0.0.0. At the commit cited here, its README said that made the agent reachable without authentication by anyone on the same network, with only settings.*, credentials.* and llm.discoverModels pinned to loopback. Current harness builds, checked 10 October 2026, authenticate the browser session with the launch token in the printed URL and a signed session cookie, but the login surface is then exposed on every interface, and a plain HTTP URL sends that token in cleartext. Use an SSH port forward, ssh -N -L 3080:127.0.0.1:3080 you@your-server, or put the port behind a private network interface plus a firewall rule, and keep TLS in front of anything reachable from outside.