trick77/opencode-presets: CLI that patches opencode.json from prepared presets. Use it to add LSP overrides, MCP servers, and permission rules without hand-editing JSON. · GitHub

Ready-made presets for OpenCode: permission rules, MCP servers, LSP overrides, agent limits, TUI preferences. A small CLI merges them into your opencode.json and tui.json so you never hand-edit the JSON.

npm install -g opencode-presets

Needs Node 22+.

Fresh install? Start with the permission bundle — one command, no prompts for the everyday read-only commands, hard blocks on the destructive ones, and session sharing turned off:

opencode-presets install permissions-recommended

It is opinionated, and it is not a free pass. These are one person's defaults for everyday work, not an audited sandbox and not a security boundary. They cut prompting for read-only commands, hard-block a list of known footguns, and set share to "disabled" — that is the whole claim. Anything not listed still falls through to a prompt you have to read; the deny rules miss env-prefixed and sh -c-wrapped invocations; and the presets you add on top can undo them. Read the rules before trusting them, and keep deciding for yourself.

Install any row below with opencode-presets install <preset>. Presets that need something outside your opencode config say so in their description.

A preset whose header is @include lines is a bundle: a list of other presets, with no rules of its own. Two ship: opencode-planify-german, and permissions-recommended, which is all of this:

permissions-recommended
  permissions-shell-safe           22 keys   ls, cat, grep, rg, jq, ps        allow
  permissions-git-safe             40 keys   status, diff, log, blame, fetch  allow
  permissions-toolchain-info       60 keys   node -v, python -V, mvn -v       allow
  permissions-container-info       52 keys   docker/podman ps, logs, inspect  allow
  permissions-deny-destructive     34 keys   sudo, dd, mkfs, rm -rf /, -f     deny
  permissions-deny-cluster-write   64 keys   oc/kubectl/helm delete, exec     deny
  privacy-share-disabled            1 key    share                            disabled

The order is part of the definition: opencode is last-match-wins and merge appends new keys at the end, so the denies have to land after every allow.

opencode-presets install permissions-recommended
opencode-presets install permissions-recommended permissions-build-tools

Reset first? Only if you already have hand-written permission.bash rules. merge never overwrites, so any rule of yours with the same pattern string as a preset's keeps that rule out — and when the casualty is a deny, a guardrail you think you installed is absent. Wiping the path first guarantees every rule lands:

jq '.permission.bash, .agent' ~/.config/opencode/opencode.json   # see what you'd lose
opencode-presets install --reset permission.bash permissions-recommended

That deletes every hand-written rule at permission.bash — a backup is written first, and nothing else in the config is touched. On a config with no bash rules of your own it changes nothing, so skip it. Either way the installer names any deny that was kept out, so you can start with a plain install and only reset if it complains.

remove expands a bundle the same way. There is no per-preset ownership tracking, so removing it also clears keys an earlier standalone install of a member wrote — the confirmation lists every preset first.

Not in the bundle, on purpose:

• permissions-build-tools — runs project-defined code, and python -c "…" / node -e "…" execute commands opencode never sees as shell commands, so no deny rule can match them.
• permissions-cluster-info — read access to whichever cluster you are logged into, production included. Worth an explicit decision.
• permissions-webfetch-ask — adds friction; wrong for a defaults bundle.

A denied command is rejected outright — no prompt, and --auto only auto-approves what is not explicitly denied. That is the one tier a habit of clicking "allow" cannot defeat. Compound commands are split before matching, so cd /x && oc delete pod y is caught too.

It is a guardrail, not a security boundary: env-prefixed (KUBECONFIG=x oc delete …), sh -c "…"-wrapped and aliased invocations slip through.

Every shipped deny pattern is disjoint from every shipped allow pattern (enforced by a test), so install order does not matter among these presets. A broad hand-written rule of your own, like "oc *": "allow", still wins if it was written after the deny — install the deny presets last.

When one of your rules keeps a deny out, the installer says so instead of letting the guardrail go missing quietly:

• permissions-deny-destructive — added 27, preserved 3
  ⚠ "sudo *" is already "ask" in your config — the deny was NOT applied
  ⚠ "rm -rf /" is already "allow" in your config — the deny was NOT applied

To apply those denies, pick one:
  1. delete the listed keys from permission.bash in ~/.config/opencode/opencode.json,
     then re-run: opencode-presets install permissions-deny-destructive
  2. wipe the whole path and reinstall from scratch:
     opencode-presets install --reset permission.bash permissions-deny-destructive
     this also deletes any other hand-written rules at that path
  3. keep your rule deliberately — nothing to do, but the guardrail is off

It also warns before you confirm if any agent sets its own permission rules: agent.<name>.permission is evaluated after the global rules and wins, so global denies do nothing for that agent.

Upgrading permissions-container-info to 0.2.0 does not revoke oc access an earlier install already granted — merge never removes keys, and 0.2.0 no longer lists the oc rules. Clear them with opencode-presets remove permissions-cluster-info.

Install multiple at once:

opencode-presets install jdtls-lombok jdtls-clean-workspace

Presets whose path uses a prompt (like mcp-http) can't be removed with remove — use reset instead:

opencode-presets reset mcp.openrag-tom

Pricing takes no configuration of its own. plugin-litellm-pricing reads each model's cost, limits and capabilities from the proxy provider-litellm already points at, so the base URL and key are the whole setup. The numbers are LiteLLM's own resolved ones, your config-level model_info overrides included, so what opencode displays is what the gateway bills. A model the proxy reports no cost for is injected without a cost block rather than with a wrong one, and the startup log names it:

grep litellm-pricing ~/.local/share/opencode/log/opencode.log

provider-litellm is a replace preset, so re-running it rewrites the whole provider.litellm block and re-prompts for the base URL and key as well — have the key to hand. Coming from an earlier install, re-run it once: that is what clears the now-unused options.catalogURL out of your config.

opencode-presets list                              # what's available
opencode-presets install jdtls-lombok             # apply one preset by name
opencode-presets install jdtls-lombok permissions-git-safe
opencode-presets remove jdtls-lombok              # undo a preset
opencode-presets install --reset permission ./presets/foo.conf  # wipe then install
opencode-presets reset permission                 # wipe a section outright
opencode-presets validate                         # check opencode.json and tui.json

Bare names are resolved through the preset search path (see "Where presets are found" below). You can always pass an explicit path instead, e.g. install ./presets/jdtls-lombok.conf.

Every change shows a diff and asks before touching anything. A backup is written to ~/.cache/opencode-presets/backups/ before each write — no auto-pruning, so they pile up.

validate checks the configured opencode.json and tui.json against OpenCode's current schemas. Use validate config, validate tui, or validate all to choose targets. Missing files are skipped in all mode; invalid files print labeled where, what, and detail lines and exit nonzero.

Presets with @prompt directives normally ask interactively. To drive them from a script (or just paste a one-liner from a wiki), pre-fill any prompt with --set NAME=VALUE:

opencode-presets install mcp-http \
  --set name=openrag \
  --set url=https://openrag.example.internal/mcp \
  --set headerName=X-Bitbucket-Token \
  --set 'headerValue=raw-token-here'

Quote values that contain shell metacharacters ($, !, *, backticks, spaces, etc.) with single quotes — otherwise the shell expands them before opencode-presets ever sees the value. A Bitbucket PAT that starts with $ will silently turn into an empty string without quoting.

For secrets, prefer --set-env NAME=ENV_VAR. The CLI reads the value from the named environment variable at install time, so the token never appears in shell history or process listings:

export BITBUCKET_TOKEN=…
opencode-presets install mcp-http \
  --set name=openrag \
  --set url=https://openrag.example.internal/mcp \
  --set headerName=X-Bitbucket-Token \
  --set-env headerValue=BITBUCKET_TOKEN

--set / --set-env apply to a single preset per invocation — run the command once per preset rather than bundling several with shared flags. This keeps the wiring obvious ("this --set goes to that preset") and avoids surprise: in a non-TTY shell script, a bundled install would happily fill the first preset's prompts and then hang on a readline for the next.

• replace — the preset owns the value at @path. Re-installing overwrites whatever's there.
• merge — the preset's keys are added; existing keys (yours or someone else's) are never overwritten. Use this for permission rules so user edits stick around.
• append — the preset's array entries are appended if missing; existing array entries are preserved. Use this for shared arrays like plugin. An entry of the form name@version supersedes every existing entry naming the same package, in place: bumping a pinned plugin replaces the old pin instead of leaving both in the array for opencode to load twice. A config that already stacked several bumps collapses to the newest one on the next install.

Re-installing is always safe: a no-op produces no backup and no write.

Plugin changes are loaded by opencode at startup. After installing a plugin preset such as plugin-superpowers, quit and restart opencode.

opencode-presets list searches dirs in this order:

1. Anything in $OPENCODE_PRESETS_PATH (colon-separated).
2. ./presets/ relative to your current directory (honoured when it exists).
3. The shipped presets/ baked into the tool.

Earlier dirs win on name collision; the lower one is still listed but flagged shadowed. Pass a positional arg (opencode-presets list ~/some/dir) to scan exactly one dir instead.

For a team or cross-machine setup, keep your presets in their own git repo (not in ~/.config, not inside the cloned tool). Point the env var at it:

# ~/.zshrc or similar
export OPENCODE_PRESETS_PATH="$HOME/work/team-opencode-presets"

Multiple repos? Colon-separate them, highest priority first:

export OPENCODE_PRESETS_PATH="$HOME/personal-presets:$HOME/work/team-presets"

For ad-hoc presets you don't want to put in a repo and don't need on other machines, drop them in ./presets/ from wherever you run the tool, or use OPENCODE_PRESETS_PATH.

OPENCODE_CONFIG=/path/to/other-opencode.json opencode-presets install ...
OPENCODE_PRESETS_CACHE=/some/cache opencode-presets install ...

TUI presets target ~/.config/opencode/tui.json by default. Override that path with OPENCODE_TUI_CONFIG:

OPENCODE_TUI_CONFIG=/path/to/tui.json opencode-presets install tui-disable-mouse

Plain JSONC with a header. Drop into one of the dirs above, or pass an absolute path.

// @name: my-preset
// @description: one or two sentences on what this fixes / sets up.
// @author: you <you@example.com>
// @version: 1.0.0
// @target: config
// @path: some.dotted.path
// @mode: merge
{ "key": "value" }

@target is optional and defaults to config, which writes opencode.json. Use @target: tui for TUI presets that write tui.json. A single install or remove operation cannot mix config and tui presets; run separate commands for those.

@fetch: <url> -> <dest> [sha256=hex] downloads to the cache. @prompt: name | text|secret|dir | help | default | setup collects input at install time; everything after the type is optional. Both repeatable. Reference fetched files as {{cache}}/<name> and prompt values as {{prompt:<name>}} in the body or @path. Leave the default field empty to skip it — name | dir | help | | <setup> has a setup hint and no default. Only the first four | are structural, so a setup hint may contain pipes of its own.

A dir prompt is checked before anything is written: the answer must be an absolute path (a leading ~ is expanded, since prompt input never sees a shell) to a directory that exists, or the install stops with an error and leaves your config alone. The resolved path is what gets written, so the entry means the same thing from any directory. --set values are checked exactly like typed ones.

The setup field is what to do to make that directory exist. It is shown in the install summary, before you are asked anything, and again if the answer fails its check:

error: prompt "skillsDir": /tmp/dd does not exist
  clone it first: git clone https://github.com/cathrynlavery/diagram-design, then answer with the skills/ dir inside that clone
  nothing was written.

@requires-bin: <name> | <setup> names an executable the preset needs on PATH, and the command that installs it. Repeatable. The check runs before the install summary, so a missing binary refuses without ever asking you to confirm; remove never checks, so you can always take one back out. It is a name resolved against PATH, not a path — the parser rejects /usr/local/bin/foo. The setup half is optional but omitting it means the refusal cannot tell you what to run:

// @requires-bin: jq | brew install jq

error: mcp-example requires "jq" on PATH.
  brew install jq
  then run this again — nothing was written.

@pins: <name> <version> records a third-party artifact the preset installs at an exact version — the npm package behind an mcp command, a plugin spec, a @fetched jar. Optional and repeatable. It's shown on the install confirmation and in list -l, so you can see what a preset drags in before saying yes:

// @pins: @playwright/mcp 0.0.79

The version string must also appear in the body or @fetch line it describes; a test enforces that, so a bump can't land on one side only.

See the existing presets/*.conf for working examples.