You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
feat(plugin): Ship the Taskade agent plugin with a front-door skill, six outcome skills, and manifests for Claude Code, Codex, Copilot and Gemini - #105
This PR ships the taskade agent plugin from this repository. One install connects an AI tool to the hosted Taskade MCP server (https://www.taskade.com/mcp) and adds 7 skills:
taskade: the front door. It names the hosted server, the sign-in, the jobs and the tools for each job, the limits, the link shapes and the safety rules.
getting-started, import-spreadsheet, fix-my-app, app-report, backup-my-app, plan-new-app: one skill for each customer job.
The same folder installs in Claude Code, Codex, GitHub Copilot CLI, VS Code agent plugins, Gemini CLI and Cursor. skills/taskade/SKILL.md is the one source text that the Taskade backend will also serve at https://www.taskade.com/SKILL.md (taskade/taskcade PR T1 embeds the same bytes).
Why
143 unique users opened the OAuth consent page for an MCP client from 2026-08-01 to 2026-10-22 (loopback 61, claude.ai 53, ChatGPT 17, Cursor web 5). Monthly: Jun 29, Jul 33, Aug 35, Sep 64, Oct 1-22 55.
14.0% of those users started a checkout within 14 days, against 2.7% for all signups.
Views of /docs/mcp* fell from 1,279 (Aug) to 159 (Oct 1-22). Agents now find tools through plugin marketplaces and skills, not through docs pages.
Today a connected agent gets the hosted tools and no guide to the jobs. The skills tell it which tools to call, in which order, and what to check before it reports back.
What changes
Files
Change
skills/*/SKILL.md (7)
Front door plus 6 outcome skills. Every tool name is a hosted tool, and every behavior claim is checked against taskade/taskcade master (list in the spec). Each job skill has a one-line link list, and the limits live in the front door.
skills/*/references/*.md (3)
workspace-model.md (parts, roles, what MCP can and cannot do), custom-fields.md (which field types set_task_field_value writes, and the refusal messages), eve-prompts.md (how to write a build prompt for Taskade EVE).
Portable Agent Plugins 1.0.0 manifest and MCP config (type: "streamable-http"). Copilot CLI, VS Code and Codex read these. The Codex listing fields sit in extensions["com.openai"].interface, as the OpenAI package guide shows ("When extensions.com.openai is an object, it replaces the entire .codex-plugin/plugin.json overlay").
.agents/plugins/marketplace.json, assets/logo.png
The Codex marketplace taskade, and the logo. The Codex interface (in the root plugin.json) follows the OpenAI submission guide: a 30-character subtitle ("Projects, agents and apps"), websiteURL, supportURL, privacyPolicyURL, termsOfServiceURL, 3 starter prompts, and the logo and composer icon as a relative path to the committed 512 px icon.
gemini-extension.json
Gemini CLI extension (httpUrl). Gemini loads skills/ with no key.
The name changes from taskade-mcp to taskade, the same name on every host (owner decision D5: no Cursor listing exists yet). Version 0.2.0, a relative logo, and a description with the plan rule. Cursor finds skills/ and mcp.json at the plugin root with no key, so the manifest drops mcpServers and the duplicate .cursor-plugin/mcp.json (its type: "http" failed its own declared schema).
README.md
The Claude plugin directory shows the README as the listing, so the first screen says what Taskade and Taskade Genesis are, then the agent plugin: what it does, the paid plan rule, what stays in Taskade, and one install table for Claude Code, Codex, GitHub Copilot CLI, Gemini CLI, VS Code, Cursor and any agent. The hosted server quick start follows, with a new ChatGPT section. The hosted tools list adds get_flow_runs and run_flow and carries no count (the kill switch USER_MCP_RUN_FLOW_DISABLED removes 2 tools). The demo table no longer says that MCP builds automations. QA: every link checked, every internal anchor resolves, the 62 local tool names match packages/server/src, the Cursor and VS Code deeplinks point at the hosted URL. Removed claims with no source: "Trusted by 3M, Nike, Tesla, Netflix, Airbnb, Disney, Adobe", "Rated 4.8/5 across 9,300+ reviews", "700+ templates", "the only MCP server", and a "task history" read that no tool does. New text has no em-dash. Old lines that this PR does not change keep their punctuation.
No package code and no server.json change, so this PR has no changeset (the same as #86 and #91).
plugin.json and mcp.json against the agent-plugins.org 1.0.0 schemas (ajv, draft 2020-12)
both valid. On main, .cursor-plugin/mcp.json failed its declared schema (type: "http").
Headless load: claude -p --plugin-dir . --model haiku "List only the skill names that start with 'taskade:' ..."
lists all 7: taskade:app-report, taskade:backup-my-app, taskade:fix-my-app, taskade:getting-started, taskade:import-spreadsheet, taskade:plan-new-app, taskade:taskade
yarn install --frozen-lockfile && yarn build && yarn lint && yarn test (the CI steps)
build 0, lint 0, test 0 (4 files, 19 tests). git diff --exit-code after the build: no tracked change.
Dash scan of every new file and every added README line
no em-dash or en-dash
README links: curl -L on all 45 URLs
200, except: vscode.dev/redirect 302 (opens VS Code), www.taskade.com/mcp 401 (OAuth start, by design), claude.ai and npmjs.com 403 (bot block), /api/v1 and /api/v2 base URLs in code text
README internal anchors (GitHub slug rules)
all 17 resolve
The npm packages, server.json and the Docker image do not change.
Owner checklist: repository settings (a PR cannot change these)
Setting
Today
Proposal
About > Description
"🤖 Taskade MCP · Official MCP server and OpenAPI to MCP codegen. Build AI agent tools from any OpenAPI API and connect to Claude, Cursor, and more."
Revert this PR. The hosted server does not depend on any file in it. The Cursor plugin goes back to taskade-mcp 0.1.2. No directory lists it yet, so the rename breaks no install.
QA steps
Step
Action
Expected
1
claude plugin marketplace add taskade/mcp, then claude plugin install taskade@taskade
The plugin installs with no error.
2
In Claude Code, run /mcp, select taskade, sign in
The Taskade sign-in page opens, and taskade shows as connected.
3
Ask "What can you do in Taskade?"
The agent loads taskade:getting-started, calls list_spaces, and shows a table of workspaces with app, projects, agents and automations.
4
Ask "Import this CSV of leads into my Leads project" with a small CSV
The agent shows the column mapping and a dry run, and writes only after a "yes".
5
Ask "Give me a status report of my workspace"
A one-screen report with counts, and an automation table from get_flow_runs.
Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.
This PR includes no changesets
When changesets are added to this PR, you'll see the packages that this PR includes changesets for and the associated semver types
This PR packages the hosted Taskade MCP integration as a multi-host agent plugin, adding guided workspace workflows without changing the npm packages or server code.
Changes:
Adds a front-door skill, six outcome skills, and supporting references.
Adds plugin manifests and marketplaces for supported AI tools, and updates Cursor metadata.
Documents installation options and expands the hosted tool list to 50.
File
Description
skills/taskade/SKILL.md
Introduces connection guidance, job routing, and safety rules.
skills/plan-new-app/SKILL.md
Guides app planning and handoff to Taskade EVE.
skills/plan-new-app/references/eve-prompts.md
Provides build-prompt guidance.
skills/import-spreadsheet/SKILL.md
Defines mapping, duplicate checks, and staged imports.
app-report calls get_flow_runs without limit, so the tool returns 10 runs, and the "20 or more" rule never fires (bug due to DEFAULT_LIMIT = 10 in get-flow-runs.ts, taskade/taskcade master)
-`list_fields` to know which fields hold status, amount, owner and date.
-`list_automations`, then `get_flow_runs` for each automation, to count runs and failures in the period. It returns at most the 20 newest runs. If all 20 are in the period, write "20 or more". It needs edit rights on automations: if it is refused, leave out the automation table and say why.
- Read-only. This skill never writes to Taskade unless the user asks to save the report.
app-report offers list_tasks as an alternative to read_project, but list_tasks returns only id, text, completed and parentId, so the report has no status, amount, owner or date fields (bug due to list-tasks.ts)
-`read_project` (or `list_tasks`) for each project in scope.
-`list_fields` to know which fields hold status, amount, owner and date.
workspace-model.md says MCP can read notes. No read tool returns a task note: read_project renders custom fields and Date and Assignee addons only, and list_tasks has no note (bug due to stringifyContentsToJSX.ts)
| Read and change projects, rows, fields, notes, dates and assignees | Add or change custom fields and their options |
| Create agents, edit their instructions, link knowledge, ask an agent | Build a new app from a prompt (Taskade EVE) |
custom-fields.md says "Never import passwords", then its closing line offers to put the value of every column of this kind in the task note, which includes Password (contradiction inside the reference)
| Media | Files and images | Upload in Taskade. A file URL can go in the note. |
| Password | A hashed secret | Never import passwords. Leave the column out. |
| Time Tracking | Timers and time entries | Leave it out |
| Yes or no (`boolean`), person (`Assign`), and system fields (`Timestamp`, `TimestampAt`, `TimestampAuthor`) | A flag, a person, or a time that Taskade records | Leave it out, or put the value in the note |
For each column of this kind, give the user three choices: leave it empty, put the value in the note, or ask Taskade EVE in that workspace to fill the field from the note after the import.
"Import this spreadsheet of leads into my Taskade CRM and skip duplicates",
"Give me a status report of my Taskade workspace for this week",
"My Taskade app shows a blank page after login. Fix it and show me what changed",
"Back up my Taskade app to this folder"
]
The front-door description says the agent can "Read and change" automations. MCP can list, read the runs of, and run an automation, and cannot create or edit one (the skill rules in this PR say so). The front door has no limits line.
description: Use Taskade from any AI agent. Read and change a user's Taskade workspace (projects, tasks, AI agents, automations and Taskade Genesis apps) through the hosted Taskade MCP server. Use when the user mentions Taskade, a Taskade Genesis app, or asks to build an app, deploy an AI agent, automate a workflow, import data into Taskade, fix or back up a Taskade app, or report on Taskade projects.
---
backup-my-app walks inspect_space only. Each mount lists at most the 100 newest projects, agents or automations (SPACE_ENTITY_LIST_LIMIT = 100), and .taskproj files have no task notes, so a large workspace is copied in part with no report of the gap
…in to taskade, and lead the README with the plugin
- app-report: read_project only (list_tasks has no field values), get_flow_runs with limit 20, and a capped failure count is never called complete.
- getting-started, workspace-model: role refusals, notes are write-only over MCP, Workspace DNA wording.
- import-spreadsheet, custom-fields: no password column in notes, task date mapping, Taskade EVE fills only Date fields.
- backup-my-app: page past the 100-entry folder listing, and say which data the .taskproj files leave out.
- fix-my-app: app/AGENTS.md only when it exists, plain modals.
- taskade (front door): automations are read and run only, and a limits line.
- Cursor plugin: name taskade, version 0.2.0, relative logo, default mcp.json discovery from the plugin root.
- Codex overlay: 30-character subtitle, support, privacy and terms URLs, three starter prompts.
- README: the Agent plugin section leads, with the plan rule and what stays in Taskade.
Impact. A customer can install Taskade into their AI tool with one command, and the agent then knows how to do real Taskade jobs: import a spreadsheet, fix a Genesis app, report on projects and automations, back up an app, and plan a new app for Taskade EVE. Today the agent gets the hosted tools and no guide, so it guesses the order and the limits. The skills carry the verified limits (20 tasks a call, one-line notes, Select takes the option id, 20 runs a read) and the safety rules (ask before each app write, say before a run that it costs credits). No server or package code changes.
Before
Customer in Claude Code / Codex / Copilot / Gemini / Cursor
|
| finds the README, copies a JSON snippet for the MCP server
v
the hosted Taskade tools, no guide
|
| "import my leads CSV"
v
agent guesses: writes rows one by one, sends the Select label
(refused), puts multi-line notes (refused), no duplicate check,
no dry run, reports "done" without a read-back
After
Customer
|
| claude plugin marketplace add taskade/mcp
| claude plugin install taskade@taskade (or codex / copilot / gemini / Cursor)
v
/mcp > taskade > Taskade sign-in (OAuth, paid plan)
|
v
"import my leads CSV"
|
v
skill taskade:import-spreadsheet
1. read file, show columns + 3 rows
2. list_fields + read_project (existing rows, option ids)
3. mapping table, one question
4. dry run: new / duplicate / problem rows -> "yes"?
5. first 50 rows -> customer checks in the app -> the rest, 20 per call
6. read back, report added / skipped / failed + local log file
Jobs it hands to Taskade EVE (cannot do through MCP):
create or edit an automation, build or publish an app, set app access
-> the skill writes the prompt, the customer pastes it into Taskade EVE
README first screen (it is also the Claude plugin listing)
Before:
# Taskade MCP Server
Connect Taskade to Claude, Cursor, VS Code, Windsurf ...
[Add to Cursor] [Install in VS Code] [Connect in Claude]
Hosted server: https://www.taskade.com/mcp
This repo holds two packages: MCP Server (stdio), OpenAPI Codegen
Contents ... Demo ... Quick Start: hosted server ... Hosted tools (48)
After:
# Taskade MCP Server and Agent Plugin
Taskade = AI-native workspace for projects, AI agents, automations.
Taskade Genesis builds a running app from one prompt.
Connect to Claude, Claude Code, ChatGPT, Codex, Cursor, VS Code, Copilot, Gemini CLI ...
[Add to Cursor] [Install in VS Code] [Connect in Claude] (same hosted URL)
## Agent plugin for Claude Code, Codex, Copilot, Gemini CLI and Cursor
Plan: every paid plan | What it does | What stays in Taskade
| Host | Install | one table: Claude Code, Codex, Copilot CLI, Gemini CLI,
VS Code (2 steps), Cursor, any agent (npx skills add ...)
Contents ... Demo
## Taskade MCP server: quick start Claude, Claude Code, ChatGPT (new), Cursor, VS Code, Windsurf, API token
## Hosted tools + get_flow_runs, run_flow (no count: a kill switch removes 2)
... local stdio package (62 tools) lower down, for developers
Also removed: unsourced claims (customer logos, review count, template count, "the only MCP server") and the demo line that said MCP builds automations. New text has no em-dash. Every link and anchor checked.
QA
Step
Action
Expected
Where
1
claude plugin marketplace add taskade/mcp then claude plugin install taskade@taskade (after merge; before merge use claude --plugin-dir <checkout of this branch>)
Plugin installs, no error
Terminal
2
In Claude Code run /mcp, pick taskade, sign in with a paid test account
"Taskade" with the subtitle "Projects, agents and apps", 7 skills, taskade MCP server
Codex CLI
10
Open the PR branch README on GitHub
First screen: what Taskade is, the plugin, the plan rule, then one install table. Contents links and the Add to Cursor / Install in VS Code buttons work
The reason will be displayed to describe this comment to others. Learn more.
🔵 Needs a closer look
Cross-client installation, authentication, and hosted-tool contracts still need end-to-end human validation beyond the supplied Claude loading and schema checks.
…d check every link, tool name and claim
- First screen: what Taskade and Taskade Genesis are, the plugin install, any-agent skills, then the hosted server for each client (ChatGPT added).
- Keyword headings for the plugin and the hosted server, with every internal anchor checked.
- Remove claims with no source (review counts, customer logos, a template count, "the only MCP server") and every em-dash.
- Privacy: the plugin runs no code of its own.
- Codex and marketplace descriptions name the paid plan and what stays in Taskade. The front door reads app/AGENTS.md only when it exists.
The reason will be displayed to describe this comment to others. Learn more.
🔵 Needs a closer look
Cross-client loading and hosted-tool contracts still need human validation, alongside the report-pagination and backup-destination corrections.
No open issues 🎉
2 previously missed issues
In code that hasn't changed since last review
Require pagination through all workspace, project, and automation lists
skills/app-report/SKILL.md:14
The report workflow does not require paging through workspace, project or automation lists. skills/getting-started/SKILL.md:14,23 documents pagination and the 100-entry page size for projects and automations. Following only the first page can silently exclude projects and automations from a workspace report, including automations with failed runs. Require exhausting these lists before calculating the report; this is separate from the explicitly handled 20-run history limit.
Prevent stale files and snapshot overwrites in backup destinations
skills/backup-my-app/SKILL.md:15
Two backups of the same workspace on the same day use the same default folder. Step 2 copies the current files but does not remove files deleted from Taskade between backups, so reusing that folder can leave stale files alongside current ones and overwrite the earlier snapshot. Require an empty destination and choose a unique folder when the proposed destination already contains files. Apply the same check to user-supplied destinations.
The hint appears only when every space on the last page is hidden for the plan (list-spaces.ts in taskade/taskcade). The skills now say that a Free workspace does not appear in the list, and to show the hint when the response has it.
…ing into the portable manifest
- README: one install table for every host, no future SKILL.md line, no hosted tool count, automations described as list, run and read runs. Old lines keep their punctuation.
- Skills: one Links line in each job skill. getting-started drops the limits that the front door states.
- plugin.json: the Codex listing fields move to extensions.com.openai.interface, and .codex-plugin/plugin.json is removed (the OpenAI package guide: the inline object replaces the overlay).
The reason will be displayed to describe this comment to others. Learn more.
🔵 Needs a closer look
Backup destination handling needs correction, and hosted-tool contracts and cross-client installation and authentication still require independent verification.
No open issues 🎉
1 previously missed issue
In code that hasn't changed since last review
Use a fresh backup snapshot directory for every run
skills/backup-my-app/SKILL.md:15
Use a fresh snapshot directory for each backup. Two backups of the same workspace on the same day select the same default path, and step 2 only copies files that currently exist. If a file or project was deleted between runs, copying into the previous directory leaves its old copy behind, mixing current and stale contents. Require a new, empty destination and choose an unused suffix when the path already exists, including for user-specified destinations.
The reason will be displayed to describe this comment to others. Learn more.
🔵 Needs a closer look
Cross-client installation, authentication, and hosted-write workflows still need end-to-end human validation beyond the supplied Claude loading checks.
No open issues 🎉
1 previously missed issue
In code that hasn't changed since last review
Clarify that skills install does not register the hosted MCP server
README.md:44
npx skills add installs the skill files, but does not register the hosted MCP server. This row appears in the plugin installation table, so users can reasonably expect the command to connect Taskade as well. Label it as a skills-only installation and include the separate remote-server setup and sign-in step so the agent has the tools these skills require.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This PR ships the
taskadeagent plugin from this repository. One install connects an AI tool to the hosted Taskade MCP server (https://www.taskade.com/mcp) and adds 7 skills:taskade: the front door. It names the hosted server, the sign-in, the jobs and the tools for each job, the limits, the link shapes and the safety rules.getting-started,import-spreadsheet,fix-my-app,app-report,backup-my-app,plan-new-app: one skill for each customer job.The same folder installs in Claude Code, Codex, GitHub Copilot CLI, VS Code agent plugins, Gemini CLI and Cursor.
skills/taskade/SKILL.mdis the one source text that the Taskade backend will also serve athttps://www.taskade.com/SKILL.md(taskade/taskcade PR T1 embeds the same bytes).Why
/docs/mcp*fell from 1,279 (Aug) to 159 (Oct 1-22). Agents now find tools through plugin marketplaces and skills, not through docs pages.What changes
skills/*/SKILL.md(7)skills/*/references/*.md(3)workspace-model.md(parts, roles, what MCP can and cannot do),custom-fields.md(which field typesset_task_field_valuewrites, and the refusal messages),eve-prompts.md(how to write a build prompt for Taskade EVE)..claude-plugin/plugin.json,.claude-plugin/marketplace.json,.mcp.jsontaskade.plugin.json,mcp.jsontype: "streamable-http"). Copilot CLI, VS Code and Codex read these. The Codex listing fields sit inextensions["com.openai"].interface, as the OpenAI package guide shows ("When extensions.com.openai is an object, it replaces the entire .codex-plugin/plugin.json overlay")..agents/plugins/marketplace.json,assets/logo.pngtaskade, and the logo. The Codexinterface(in the rootplugin.json) follows the OpenAI submission guide: a 30-character subtitle ("Projects, agents and apps"),websiteURL,supportURL,privacyPolicyURL,termsOfServiceURL, 3 starter prompts, and the logo and composer icon as a relative path to the committed 512 px icon.gemini-extension.jsonhttpUrl). Gemini loadsskills/with no key..cursor-plugin/plugin.json,.cursor-plugin/mcp.json(removed)taskade-mcptotaskade, the same name on every host (owner decision D5: no Cursor listing exists yet). Version 0.2.0, a relativelogo, and a description with the plan rule. Cursor findsskills/andmcp.jsonat the plugin root with no key, so the manifest dropsmcpServersand the duplicate.cursor-plugin/mcp.json(itstype: "http"failed its own declared schema).README.mdget_flow_runsandrun_flowand carries no count (the kill switchUSER_MCP_RUN_FLOW_DISABLEDremoves 2 tools). The demo table no longer says that MCP builds automations. QA: every link checked, every internal anchor resolves, the 62 local tool names matchpackages/server/src, the Cursor and VS Code deeplinks point at the hosted URL. Removed claims with no source: "Trusted by 3M, Nike, Tesla, Netflix, Airbnb, Disney, Adobe", "Rated 4.8/5 across 9,300+ reviews", "700+ templates", "the only MCP server", and a "task history" read that no tool does. New text has no em-dash. Old lines that this PR does not change keep their punctuation.No package code and no
server.jsonchange, so this PR has no changeset (the same as #86 and #91).Manifest sources: Agent Plugins 1.0.0 spec and schemas (https://agent-plugins.org), OpenAI package and submission guides (https://developers.openai.com/plugins/build/plugins, https://developers.openai.com/plugins/deploy/submission), Copilot CLI plugin reference (https://docs.github.com/en/copilot/reference/copilot-cli-reference/cli-plugin-reference), VS Code agent plugin reference (https://code.visualstudio.com/docs/agents/reference/agent-plugin-reference), Cursor plugin reference (https://cursor.com/docs/reference/plugins), Gemini CLI extension docs (https://geminicli.com/docs/extensions/writing-extensions/).
Zero-regression evidence
claude plugin validate .✔ Validation passed(marketplace)claude plugin validate .claude-plugin/plugin.json✔ Validation passedplugin.jsonandmcp.jsonagainst the agent-plugins.org 1.0.0 schemas (ajv, draft 2020-12)main,.cursor-plugin/mcp.jsonfailed its declared schema (type: "http").claude -p --plugin-dir . --model haiku "List only the skill names that start with 'taskade:' ..."taskade:app-report,taskade:backup-my-app,taskade:fix-my-app,taskade:getting-started,taskade:import-spreadsheet,taskade:plan-new-app,taskade:taskadeyarn install --frozen-lockfile && yarn build && yarn lint && yarn test(the CI steps)git diff --exit-codeafter the build: no tracked change.curl -Lon all 45 URLsvscode.dev/redirect302 (opens VS Code),www.taskade.com/mcp401 (OAuth start, by design),claude.aiandnpmjs.com403 (bot block),/api/v1and/api/v2base URLs in code textThe npm packages,
server.jsonand the Docker image do not change.Owner checklist: repository settings (a PR cannot change these)
ai-agent,autonomous-agents,model-context-protocol-servers,openapi3,workflowswithclaude-code-plugin,agent-skills,codex-plugin,gemini-cli-extension,cursor-pluginRollback
Revert this PR. The hosted server does not depend on any file in it. The Cursor plugin goes back to
taskade-mcp0.1.2. No directory lists it yet, so the rename breaks no install.QA steps
claude plugin marketplace add taskade/mcp, thenclaude plugin install taskade@taskade/mcp, selecttaskade, sign intaskadeshows as connected.taskade:getting-started, callslist_spaces, and shows a table of workspaces with app, projects, agents and automations.get_flow_runs.gemini extensions install https://github.com/taskade/mcp(after merge)taskadeMCP server.copilot plugin install taskade/mcp(after merge)taskadeMCP server.codex plugin marketplace add taskade/mcp, thencodex plugin add taskade@taskade(after merge)taskadeMCP server.Refs taskade/taskcade#32572