brightbean-studio
Open-source, self-hostable social media management platform. Schedule, publish, and manage content across 10+ platforms from a single dashboard. Free alternative to Buffer, Sendible, and SocialPilot.
- 2.7k
- Python
- AGPL-3.0
- 5 d ago
Plugins for MiniMax Code: drive your real signed-in Chrome, generate images and video, and retry instead of giving up. Zero dependencies.
Repo DraconDev/mcode-plugins, read from GitHub 1 min ago. DraconDev has not joined Take My Repo.
DraconDev/mcode-plugins is a public repo on GitHub by DraconDev. Plugins for MiniMax Code: drive your real signed-in Chrome, generate images and video, and retry instead of giving up. Zero dependencies. It declares the AGPL-3.0 license on GitHub. It is written mainly in JavaScript, has 1 stars and last changed on 2026-11-07. Nobody has reviewed it on Take My Repo yet, which says nothing about its quality. Read from GitHub on 2026-11-22.
In the maintainer's words. Shown from GitHub; Take My Repo does not keep a copy.
Plugins for MiniMax Code (@minimax-ai/code) that give it
three things it does not ship with: a browser it can actually drive, image and video generation,
and a retry policy that does not give up after half a minute.
Everything here is plain Node — no node_modules, no build step, no framework. Clone it, run one
script, and the plugins are installed.
| Plugin | What it does |
|---|---|
mcode-chrome/ |
Drive the user's real, signed-in Chrome — real logins, real tabs, the tab already open. 23 tools. |
mcode-agnes/ |
Agnes AI image and video generation. Set one environment variable. 3 tools. |
mcode-codex/ |
ChatGPT-subscription (Codex) image generation and editing. Set one environment variable. 2 tools. |
mcode-retry-on-error/ |
Continue a turn that ended in an error or a rate-limited tool call, with backoff. Hook-based, no tools. |
fetch and AbortSignal.timeout. Check with node --version.schemaVersion: 1) and the hook events
these plugins use are that version's.mcode on PATH — for the install script to register the plugin.mcode-chrome only: the Pi Chrome Connector browser extension, loaded into your normal
Chrome profile. See mcode-chrome/README.md.git clone https://github.com/DraconDev/mcode-plugins.git
cd mcode-pluginsThen install one plugin, or all four:
scripts/install-plugin.sh mcode-chrome scripts/install-plugin.sh mcode-agnes scripts/install-plugin.sh mcode-codex scripts/install-plugin.sh mcode-retry-on-error
# or all of them for plugin in mcode-chrome mcode-agnes mcode-codex mcode-retry-on-error; do scripts/install-plugin.sh "$plugin" done
The script copies the plugin to ~/.minimax/plugins/<name>/, where local plugins are enabled by
presence, and registers it with the local marketplace. Verify with:
mcode plugin list
You should see each plugin as enabled. To update a plugin, re-run its install script — it replaces
the installed copy in place.
Installing a plugin does not stop you from using mcode without it. Each one is independent; install only the ones you want.
Only the two image plugins need a credential. Each one takes an environment variable, and that is all it needs — no other tool has to be installed:
| Plugin | Environment variable | What it is |
|---|---|---|
mcode-agnes |
AGNES_API_KEY or AGNES_CN_API_KEY |
your Agnes API key (the two cover the two Agnes regions) |
mcode-codex |
CODEX_IMAGE_ACCESS_TOKEN |
the ChatGPT subscription's image OAuth access token. Add CODEX_IMAGE_ACCOUNT_ID if you want to be explicit about the account. |
export AGNES_API_KEY="…" export CODEX_IMAGE_ACCESS_TOKEN="…"
Put them wherever you already keep environment variables — a shell profile, direnv, your
session manager. The plugin reads them at call time, so there is nothing to configure and no
restart to do.
If you already use pi and have signed in to Agnes or to the
Codex image provider there, these plugins will pick that credential up on their own from pi's auth
store (~/.pi/agent/auth.json) — one login, two agents, no environment variable to set. pi is
not required; the store is a convenience, not a dependency. When both are available the
environment variable wins.
The codex plugin refreshes an expired token in memory only and never writes back to pi's auth file, so it cannot corrupt another tool's state. A token you supply yourself is never refreshed — supply a fresh one when it expires.
mcode-chrome needs no credential of its own; it uses the session your Chrome extension already
has.
mcode-chrome — drive your real browserA MiniMax Code plugin that drives the user's own Chrome through a companion extension, rather than a separate automation browser. That means the agent works inside the session you are already signed in to, on the tabs you already have open.
status, authorize, revoke, tab, snapshot, find, inspect, navigate, evaluate,
click, type, fill, key, wait_for, console_messages, network_requests,
network_request, screenshot, hover, drag, tap, scroll, upload_file
Setup, the wire protocol, and the security posture are documented in
mcode-chrome/README.md. The short version: load the extension, run
mcode-chrome's install.sh, and keep that Chrome window open.
mcode-agnes — Agnes AI images and videostatus, generate_image, generate_video
Takes a prompt, optional reference images, and writes the result to disk. A savePath relative to
the workspace lands in the workspace; baseDir overrides the base. A video can be driven from
text-to-video, one keyframe, or two.
mcode-codex — ChatGPT subscription imagesstatus, generate_image
Generates from a prompt and up to five reference images, and can edit an existing picture in a
consistent style. The ChatGPT subscription has its own quota and will answer 429 under load —
which is why the retry policy below exists.
mcode-retry-on-error — do not give upTwo hooks:
PostToolUse fires the instant a tool fails, while the model is still working, and injects a
retry instruction into the tool result.Stop is the backstop for a turn that tried to end in an error — it blocks the turn and feeds
a retry instruction back to the model.Both only act on transient failures. A validation error, a permission denial, a missing file, an
authentication failure, a content filter — those are terminal, and retrying them just burns the
user's time and quota. A rate limit, a timeout, a dropped connection, a 503 — those clear on
their own, so the plugin keeps trying.
The image plugins share one retry core, shared/retry.mjs:
fetch failed. A server's
Retry-After is honoured.mcode-retry-on-error is a different mechanism for the same goal: it works at the tool and turn
level, so it also covers failures in tools that have no retry policy of their own.
MiniMax Code retries a rate-limited provider request itself, and that error never reaches the
session transcript — the Stop hook does not fire for it, because the run aborts before a turn
completes. Measured: across 7,185 assistant messages in a day's transcripts, the stopReason is
only ever toolUse or stop — never error.
So the provider layer has to be repaired in the mcode bundle itself. See Runtime repair below.
This is optional, and it patches a third-party install. MiniMax Code retries a rate-limited provider request with a hardcoded budget:
{maxRetries: 5, baseDelayMs: 1000, maxDelayMs: 30000, maxRetryElapsedMs: 120000}
Five retries means six attempts over about twenty seconds of waiting, then the turn dies. There is no way to configure it: the policy is merged in by a minified function, but the value it would merge never arrives — there is no config key and no environment variable for it anywhere in the bundle. The constant is the only lever.
scripts/apply-mcode-llm-retry-policy.mjs raises it to a budget that outlasts a quota window
rather than a momentary burst:
{maxRetries: 200, baseDelayMs: 1500, maxDelayMs: 300000, maxRetryElapsedMs: 21600000}
The first few attempts stay close together, so a rate limit that clears in seconds is caught almost immediately. The delay then doubles up to a five-minute cap, and the loop keeps going for six hours — long enough to outlast a quota that takes hours to reset. The attempt count is a backstop; the wall clock is what actually binds.
It also patches the classifier, so that a spent quota, exhausted credits and a plain 5xx are retried instead of being treated as terminal:
| Code | Meaning | Before | After |
|---|---|---|---|
| 50150 | LLM_TPM_RATE_LIMITED |
retried | retried |
| 50111 | LLM_RATE_LIMITED |
retried | retried |
| 50151 | LLM_CLUSTER_OVERLOADED |
retried | retried |
| 42212 | USAGE_LIMIT_EXCEEDED |
terminal | retried |
| 50110 | LLM_CREDITS_EXHAUSTED |
terminal | retried |
| 5xx | a server error | terminal | retried |
| 50112 | LLM_AUTH_ERROR (401/403) |
terminal | terminal |
| — | a content filter, a refusal, a bad request, an explicit abort | terminal | terminal |
An authentication failure stays terminal on purpose: a bad credential will not fix itself, and retrying it for six hours is pointless.
node scripts/apply-mcode-llm-retry-policy.mjs # patch it node scripts/apply-mcode-llm-retry-policy.mjs --check # report only, no writes node scripts/apply-mcode-llm-retry-policy.mjs --revert # back to MiniMax Code's default node scripts/verify-mcode-retry-classifier.mjs # test the installed classifier
The original chunk is kept as <chunk>.orig.bak beside it, and the script refuses to overwrite a
policy it does not recognise. mcode update replaces the bundle wholesale, so re-run the script
afterwards — that is the whole reason it is a script and not a patch file.
Restart after patching. Node reads the bundle into memory when a process starts, so a patch on disk only reaches processes started afterwards. The script prints the sessions that are still running the old code:
WARNING: 19 mcode processes are still running the old code.
PID 3804772 started 2026-10-19 12:54:09
The oldest has been up for 5h 26m.
They keep the old budget — 5 retries, about half a minute — until restarted.
Quit those sessions and start mcode again to pick up the new policy.
The same warning appears in --check mode.
MiniMax Code seeds a recoverable-delete rm shim at <dataDir>/shims/rm and puts it ahead of
/bin/rm on the agent's PATH, so every delete lands in the platform trash instead of being
unlinked. The shim is generated from a template in the mcode bundle whose shebang is hardcoded to
#!/bin/bash.
On a Nix-based system — where bash lives at /run/current-system/sw/bin/bash and /bin/bash does
not exist — the shim cannot exec at all, and every rm in agent bash fails:
/home/dracon/.minimax/shims/rm: /bin/bash: bad interpreter: No such file or directory
scripts/apply-mcode-rm-shim-fix.mjs changes the template's shebang to #!/usr/bin/env bash,
which resolves bash through PATH on both Nix-style and conventional systems, and re-seeds the live
shim so the fix applies immediately.
node scripts/apply-mcode-rm-shim-fix.mjs # patch the template and re-seed the shim node scripts/apply-mcode-rm-shim-fix.mjs --check # report only, no writes
Every plugin has a test script, and they all need nothing but Node:
node scripts/retry-test.mjs # the retry policy node scripts/workspace-test.mjs # the workspace directory resolver node scripts/mcp-server-test.mjs # the MCP stdio transport node scripts/install-script-test.mjs # the plugin installer node scripts/credential-test.mjs # credential resolution, with and without pi node scripts/sync-shared.mjs --check # shared modules in sync node mcode-chrome/scripts/server-mode-test.mjs # standalone protocol round trip node mcode-chrome/scripts/regression-test.mjs # redaction, tab groups, screenshots node mcode-agnes/scripts/path-safety-test.mjs # path containment and image limits node mcode-codex/scripts/path-safety-test.mjs # path containment and image limits node mcode-chrome/scripts/smoke-test.mjs # MCP handshake against the live bridge node mcode-agnes/scripts/smoke-test.mjs # status only, no credits spent node mcode-agnes/scripts/smoke-test.mjs --generate # also generate one real image node mcode-codex/scripts/smoke-test.mjs # status only, no credits spent node mcode-codex/scripts/smoke-test.mjs --generate # also generate one real image node mcode-retry-on-error/scripts/hook-test.mjs # the Stop hook decisions node mcode-retry-on-error/scripts/post-tool-hook-test.mjs # the PostToolUse decisions node scripts/verify-mcode-retry-classifier.mjs # the installed provider classifier
--generate spends real quota. The other two smoke tests only check the handshake and the
credential.
Each plugin has a scripts/build-manifest.mjs that derives mcp/<server>.mcp.json from the tool
definitions in server/tools.mjs, so the declared tool list can never drift from the running
server. Run all three after changing a tool:
node mcode-agnes/scripts/build-manifest.mjs node mcode-codex/scripts/build-manifest.mjs node mcode-chrome/scripts/build-manifest.mjs
CI checks that regenerating them produces no diff.
shared/ holds the single copy of the retry core, the MCP server and the workspace resolver. Each
plugin needs its own copy on disk because a plugin is installed by copying its directory somewhere
else entirely — it has to be self-contained. scripts/sync-shared.mjs keeps those copies identical:
node scripts/sync-shared.mjs # copy shared/ into every plugin that has a server/ node scripts/sync-shared.mjs --check # fail if they have drifted
Never hand-edit a plugin's copy of retry.mjs, mcp-server.mjs or workspace.mjs — edit
shared/ and re-run the sync.
MiniMax Code spawns plugin MCP servers with the home directory as their working directory and
passes no workspace path in the environment, so a relative savePath would land in ~/….
shared/workspace.mjs resolves the workspace instead — MCODE_WORKSPACE if it is set, otherwise
PWD, which is the directory mcode was launched from. Every path-taking tool also accepts an
explicit baseDir. Absolute paths always work and are preferred.
node executable with no
node_modules, so they can be copied anywhere mcode runs.*.orig.bak).The plugin does not show up in mcode plugin list.
Local plugins are enabled by presence in ~/.minimax/plugins/<name>/. Check the directory exists
and contains .minimax-plugin/plugin.json. Then re-run the install script.
A tool call fails with "Tool-result delivery seam is closed".
Something is returning a top-level {"decision":"block"} from a PostToolUse hook. That form is
correct for the Stop hook and forbidden for PostToolUse, which must use
hookSpecificOutput with additionalContext instead.
A rate limit still gives up after about half a minute.
The session was started before the bundle patch. Node loads the bundle into memory at startup.
Run node scripts/apply-mcode-llm-retry-policy.mjs --check — it lists the sessions that are still
running the old code — and restart them.
rm fails with "bad interpreter" in agent bash.
The recoverable-delete shim was generated for a system with /bin/bash. Run
node scripts/apply-mcode-rm-shim-fix.mjs.
The retry hook does not seem to fire.
Its data directory is ~/.minimax/v2/plugin-data/hooks/mcode-retry-on-error/ — not a
.plugin-data folder next to the plugin. Set "debug": true in that directory's config.json and
the heartbeat log next to it will show every invocation.
See CONTRIBUTING.md. The short version: keep it dependency-free, generate the manifests, sync the shared modules, and add a test for anything you fix.
AGPL-3.0-or-later. See LICENSE.
DraconDev/mcode-plugins is a public repo on GitHub by DraconDev. Plugins for MiniMax Code: drive your real signed-in Chrome, generate images and video, and retry instead of giving up. Zero dependencies.
It declares the AGPL-3.0 license on GitHub.
Its maker has not tagged a system, so it is unknown here. Its README says what it needs.
Open "Take it" on its page: Take My Repo writes an install plan for your machine and hands it to your AI agent (Claude Code, Codex or Cursor), which asks you before each step. Take My Repo has not tested it on any machine.
Its last change on GitHub was on 2026-11-07. GitHub counts 0 open issues and pull requests. Take My Repo does not judge from that whether it is maintained.
Nobody has reviewed it on Take My Repo yet, which says nothing about its quality.

Open-source, self-hostable social media management platform. Schedule, publish, and manage content across 10+ platforms from a single dashboard. Free alternative to Buffer, Sendible, and SocialPilot.
🔮 Platform-neutral Browser Use for AI agents: real Chrome automation with a CLI + SDKs, no lock-in, dead simple.

MR Office: open-source document, spreadsheet, presentation, PDF, handwriting and code editors. https://mroffice.ai

Browser automation plugin for OpenCode AI editor - Control Chrome/Edge with AI, automate web testing, scraping & form filling via MCP integration
World's first open-source, agentic video production system. 12 production pipelines, 100+ tools, 700+ agent skill and production-knowledge files. Turn your AI coding assistant into a full video production studio.

Long-running autonomy for Pi: draft goals, build projects from specs, run audited task queues, and recover interrupted work. Independent auditors verify completion.
Lightweight Rust daemon for controlling OBS Studio with global hotkeys on Wayland and X11
Git AI Auto Committer is a VS Code extension that automatically commits your code changes with intelligent, AI-generated messages. It monitors your coding activity and commits your work when you naturally pause, ensuring your progress is saved without interrupting your flow.
Tiles — TUI file manager with git integration, SSH, system monitor, and text editor.
Auto-continue stalled OpenCode sessions — stall recovery, todo nudging, review on completion, and AI-driven session analysis
Rust CLI & library for uploading videos to YouTube via the Data API v3. Resumable uploads, multi-channel workspaces, upload profiles, per-video metadata TOML, JSON output for automation.
This page is also available as structured data, with the same sources and dates.
/api/v1/resources/gh/DraconDev/mcode-pluginsTake My Repo has not tested this on your machine. Requirements come from the project's own documentation.