Plugin registry
Plugins for pansy, OpenCode and Pi, ranked by what npm, GitHub and their users say.
- 51N
A push when an agent's turn is done or fails, after it ran long enough to be worth one: ntfy, Bark, Gotify, Server酱, Feishu, DingTalk, WeCom, Telegram or your own webhook. Also pushes when a subscription account is signed out and needs signing in again. Agent, project, model and time only by default; the reply's text is opt-in.
0Weekly downloads0Stars0Users0Likes - 52N
Give every harness behind Magpie Nowledge Mem automatically: recall and the Context Bundle (or Working Memory) before each request, per-harness provenance, optional import of completed conversations as Mem Threads, and a real connection test in the page. A local or remote Mem; saving conversations is off by default.
0Weekly downloads0Stars0Users0Likes - 53
Recall from your PowerContext service and optionally save complete conversations. Configure the service URL first; capture is off by default.
0Weekly downloads0Stars0Users0Likes - 54
Help AI agents evolve through experience: per-turn recall and complete engine-policy reads, automatic process diaries and grounded archival with write readback. Configure an existing local Runtime and generation model first. Archival defaults on; raw capture off. Background storage happens after reply completion.
0Weekly downloads0Stars0Users0Likes - 55M
Email alerts when a provider call fails or a quota window runs out: your own SMTP account, your own words in the mail, and your own list of what to watch, with cooldowns and an hourly cap so a bad hour stays bounded — plus an optional periodic usage digest.
0Weekly downloads0Stars0Users0Likes - 56A
Claude Code and Claude Desktop on an Antigravity account: the words Antigravity refuses their system prompt for are broken up, in what Antigravity alone is sent.
0Weekly downloads0Stars0Users0Likes - 57C
Opt-in, per-provider Codex Responses history filtering after switching providers. Preserves readable content and tool-call pairing; no model probes.
0Weekly downloads0Stars0Users0Likes - 58C
Recognize and mask company names with stable session aliases, restore replies, and show local receipts and explicit Codex session approvals. Unlike Word Guard, names can be restored. Text only; rule-based detection.
0Weekly downloads0Stars0Users0Likes - 59E
Drop the end of a reply that has reasoning but no text, so Claude Code and Codex retry instead of treating an empty turn as finished. Anthropic and Responses, streamed or whole.
0Weekly downloads0Stars0Users0Likes - 60F
Thin an agent's old history while preserving a recent window and tool pairing. Optional saccade collapses old content on a new topic; a structured state block replaces the raw task message; cut boundaries move in blocks to keep provider prompt caches settling.
0Weekly downloads0Stars0Users0Likes - 61H
Send a request's history smaller: fold a run of blank lines, drop a system prompt written twice and a message repeated right after itself, and cut a past tool result longer than maxToolLength at a word boundary. The turn being sent now is left whole.
0Weekly downloads0Stars0Users0Likes - 62M
New API's model redirection: send a model under another name, by name or pattern; replies keep the name asked for.
0Weekly downloads0Stars0Users0Likes - 63O
Seven composable output styles, all off by default, with a selection panel and prompt-cache-stable injection that preserves Codex instructions. Unofficial; by manjughosa.
0Weekly downloads0Stars0Users0Likes - 64P
New API's parameter override: set, delete, move or rewrite fields of every request, by model or by condition.
0Weekly downloads0Stars0Users0Likes - 65
Maintained by the ShyPie project. The Magpie-side gate of ShyPie's privacy-first workflow for DeepSeek Harness: with DSH on DeepSeek models behind Magpie's gateway, it turns away image/file/audio parts, encoded blobs and models outside the allowlist whenever Magpie cannot mask them locally. It does not mask text (Magpie's Privacy masking does); images and files get a 403 by default unless localModels and localProviders both allow them. Set binaryPolicy: guide to withhold binary parts and ask the model to choose a local extraction tool instead; structural checks fail closed; a request above maxNodes values is refused, but text beyond scanChars is not inspected (newest first, each field capped), so long encoded tokens can be false positives.
0Weekly downloads0Stars0Users0Likes - 66R
Two routings that keep something in reserve: fill accounts up to a ceiling of their allowance, or use subscriptions before API keys.
0Weekly downloads0Stars0Users0Likes - 67S
Add your own system prompt to every request, or only some agents' or models', in each API's own place.
0Weekly downloads0Stars0Users0Likes - 68T
Take a model's leading <think>, <thinking> or <thought> block out of its reply's text, or put a Chat reply's reasoning into it.
0Weekly downloads0Stars0Users0Likes - 69W
New API's sensitive-word filter: turn away or mask words and patterns in what you send, and in replies.
0Weekly downloads0Stars0Users0Likes - 70

Agnes AI international with Chrome browser sign-in: live OpenAI-compatible text models, Token Plan and Agent Plan usage, and the USD wallet balance.
0Weekly downloads0Stars0Users0Likes - 71A
Unofficial Agnes Code desktop: reuse the desktop sign-in for its OpenAI-compatible gateway, with the live model catalog, daily credit check-in and points balance. Desktop decryption is Windows-only; a paste-token method covers the rest. Differs from opencode-agnes-auth (`agnes`), which signs in to Agnes AI international in Chrome and meters that product's plans and USD balance: this one serves the Agnes Code desktop product and its daily credits.
0Weekly downloads0Stars0Users0Likes - 72A
Unofficial provider that runs the agy CLI you installed and signed in to through the Agent Client Protocol (agy acp), serving only the models agy offers this client. Unlike Antigravity ACP, which downloads Google's ACP server and rides the Antigravity IDE's sign-in, it uses your own agy and adds a process pool, a history budget, a bridge for the client's tools, session resume and agy's usage windows.
0Weekly downloads0Stars0Users0Likes - 73A
Unofficial AStudio (iFLYTEK Astron Studio): sign in through a browser (each account gets its own pool slot), reuse the desktop sign-in, or paste a token, to serve its Responses model gateway with the live model catalog, daily login-points check-in (read from the credit ledger) and points balance with spark-point expiry. Differs from Providers Hub's hub-astudio, which signs in on a local page (SMS, password or WeChat QR) and calls chat/completions; on one account both share the same daily points pool.
0Weekly downloads0Stars0Users0Likes - 74
Unofficial AtomGit CodingPlan access for magpie and OpenCode: browser OAuth or local AtomCode CLI sign-in, signed OpenAI-compatible requests, live models, quota windows, and a read-only diagnostics page.
0Weekly downloads0Stars0Users0Likes - 75A
AutoClaw (Zhipu OpenClaw) cloud accounts: oversea web sign-in (Google/z.ai) or CN phone SMS, credit balance and daily check-in.
0Weekly downloads0Stars0Users0Likes - 76C
Meituan's 妙手 CatPaw (ai.catpaw.meituan.com): a browser web sign-in or the desktop client's saved credentials, chat through its round/event/turn conversation protocol, models and credit balance.
0Weekly downloads0Stars0Users0Likes - 77C
Cloudflare Workers AI with your API token (Workers AI:Edit): 28+ models discovered live from models.dev, the 8 paid-only models flagged, and a Neurons/day allowance row in magpie quota.
0Weekly downloads0Stars0Users0Likes - 78C
Unofficial: Huawei Cloud CodeArts (码道) coding plans and OfficeAce (AgentArts), each signed in with Huawei Cloud OAuth on its own card — GLM, DeepSeek, openPangu and OfficeAce's MaaS models, live models and credit rates, credit usage by bucket and the daily check-in. Replaces officeace-auth; accounts signed in there keep working.
0Weekly downloads0Stars0Users0Likes - 79G
GitHub Copilot on GitHub.com or GitHub Enterprise, with premium request or credit usage, remaining allowance and quota reset dates.
0Weekly downloads0Stars0Users0Likes - 80豆
Not a model API: Doubao Work has none for other clients. This drives its desktop app (or Doubao's) the way a person does, so each request is one run of Doubao's cloud agent in a new task, with the client's tools asked for as text. macOS and Windows; unofficial.
0Weekly downloads0Stars0Users0Likes - 81G
Vertex AI models and Google Cloud promotional credits, signed with your own gcloud or ADC login. Lists only the models your project can call.
0Weekly downloads0Stars0Users0Likes - 82
Gumloop with your API key and user ID: account models, streaming, tool calls and a locally calibrated credit estimate.
0Weekly downloads0Stars0Users0Likes - 83
Unofficial: Intern InkStone(书生·端砚) Token Plan: a browser sign-in reads the ink-credit balance, the 5-hour and 7-day windows and the credit packs, and creates an ink-* API key (or reuses a pasted one) to serve the plan's models for chat.
0Weekly downloads0Stars0Users0Likes - 84J
Unofficial JoyCode provider: JD account browser sign-in, live models and credit multipliers, plan credits and streaming tool calls.
0Weekly downloads0Stars0Users0Likes - 85K
Kimi desktop Work subscription through its local sign-in, with live Work models and shared membership usage. Separate from Kimi Code.
0Weekly downloads0Stars0Users0Likes - 86L
iFlytek's Loomy (loomyad.xunfei.cn): SMS sign-in, OpenAI-compatible chat with reasoning-effort folding and image markers, the live model catalog, two point wallets and the daily first-login grant.
0Weekly downloads0Stars0Users0Likes - 87

Unofficial MClaw provider: China Mobile Cloud's MClaw agent (ai.yun.139.com) with an OpenAI-compatible endpoint — the browser sign-in harvests the ticket by itself, auto-fetches the llm key and the token multipliers, serves the live model catalogue and the aitokens pool usage, and renews the ticket before it expires.
0Weekly downloads0Stars0Users0Likes - 88N
Nous Portal models, with Nous Portal's device sign-in.
0Weekly downloads0Stars0Users0Likes - 89O
OpenAI's native Decisions API as System One for Magpie classifiers and decision groups: choice, score and boolean probabilities, with standard API-key sign-in and per-question refusal handling.
0Weekly downloads0Stars0Users0Likes - 90O
Tencent Cloud OrcaTerm (遨驰终端) AI assistant models — DeepSeek, Hunyuan, GLM, Kimi — on the account's monthly question allowance. Unofficial; the README sets out the terms.
0Weekly downloads0Stars0Users0Likes - 91商
SenseTime Raccoon (xiaohuanxiong.com) models, account credits and daily check-in, with browser sign-in or Raccoon desktop's live sign-in.
0Weekly downloads0Stars0Users0Likes - 92R
River (river.ai) pay-as-you-go inference with your API key: per-account models from /v1/models, documented reasoning efforts and live prices from River's public pricing endpoint.
0Weekly downloads0Stars0Users0Likes - 93S
Existing SAP AI Core deployments with client-secret credential-file sign-in, per-resource-group GPT Responses and Claude Messages discovery, streaming and agent tools. No deployment provisioning or quota reporting; Magpie only.
0Weekly downloads0Stars0Users0Likes - 94S
SenseNova (SenseTime 日日新) Token Plan: a browser sign-in (or the console account and password) reads the credit pools, their 5-hour and 7-day windows and the expiring grants, and serves the plan's own models — the browser sign-in reuses or creates a Token Plan API key for you.
0Weekly downloads0Stars0Users0Likes - 95
Unofficial Tabbit (web.tabbit.com, and web.tabbit.ai for international accounts) provider: reuse the Tabbit Browser sign-in on this computer to serve its chat models, with the account's live catalog, cycle usage and the daily check-in. No browser automation: a turn is Tabbit's own room → run → event-stream calls. Client tool calls are not available (Tabbit's tools are its own).
0Weekly downloads0Stars0Users0Likes - 96W
Warp (the terminal) AI request credits through the Warp app's sign-in: the auto routers and every model family the account lists.
0Weekly downloads0Stars0Users0Likes - 97A
Use your terminal Agent CLIs right inside magpie, by chat. Runs a local service on 127.0.0.1 that starts when the page opens and stops with magpie; only its own page gets the token. First install ~46 MB; Claude Code isn't bundled, so install the `claude` CLI yourself.
0Weekly downloads0Stars0Users0Likes - 98A
Reads the session files of Claude Code, Codex, Copilot CLI, Kimi Code, DeepSeek Harness, Gemini CLI, Qwen Code, Pi, omp, Qoder, WorkBuddy, CodeBuddy and Grok Build on this computer (prompts, replies and tool commands; text can be turned off). Shows them in a real WebGL 3D office with eleven scenes, an isometric/2D office, or a grouped workbench with Today usage, recent conversations and approvals.
0Weekly downloads0Stars0Users0Likes - 99
An Amp usage board: requests, tokens and cache hit rate by hour, day and model, read from your Amp account and legacy local threads. Its background service syncs that usage into Magpie requests and Day card, with duplicate checks; set sync to false to only view it.
0Weekly downloads0Stars0Users0Likes - 100C
A chat page in magpie's window by Magic_zero: talk to magpie's own models with your MCP servers, instructions and skills, images and LaTeX. Runs a local service on 127.0.0.1 that only its own page can use.
0Weekly downloads0Stars0Users0Likes