Priorify.app Agent Kit
Give your coding agent a safe, shared operating context for product work.
Priorify.app · Agent Skills catalogue · MCP documentation · Releases
Priorify.app Agent Kit is the official collection of portable Agent Skills and
plugin manifests for using Priorify.app through its hosted
MCP server. One repository works with Codex, Claude Code, Cursor, GitHub
Copilot, Gemini CLI, Grok, and other Agent Skills or Agent Plugins clients.
The kit contains instructions and connection metadata only. Product data stays
in Priorify.app, authentication uses OAuth, and every agent remains limited by the
workspace, product, capability, and approval boundaries selected by the user.
Included skills
| Skill | Purpose |
|---|
priorify-product-operations | Review product context, projects, Knowledge, search and store evidence, and delivery state, then make authorized updates. |
priorify-social-planning | Review and maintain owned social calendars without pretending to publish externally. |
priorify-agent-work | Claim delegated work, report safe progress, request intervention, and return durable results. |
priorify-agent-sessions | Report what a chat is doing: start a session, report milestones, ask a person to decide and wait for the answer on the board, and end it. |
Report your chats
Paste this into CLAUDE.md, AGENTS.md or GEMINI.md in each repository so
its chats appear on the Agent activity page in Priorify.app. Replace
<product id> and <workspace id> with the ids of the product the repository
belongs to and its workspace; list_products returns both, and the product page
URL shows the product id.
## Report your chats to Priorify.app
This repository belongs to Priorify.app product `<product id>` (copy it from the product page URL or `list_products`) in workspace `<workspace id>` (`list_products` returns it, and the app URL shows it as `workspace=`). Pass both as `product_id` and `workspace_id` on every call. Priorify.app only shows what you report; it never runs or wakes you.
- When a substantial piece of work starts, call `start_agent_session` with that `product_id`, a one-line `goal`, your client name as `client_label`, the repository or folder and branch as `locator`, the link to this chat as `chat_url` if your client has one, and a `client_session_key`:
- Claude Code: the value of the environment variable `CLAUDE_CODE_SESSION_ID`.
- Another client: its own session or thread id, if it exposes one.
- Otherwise `<client>:<repository>:<branch>:<yyyy-mm-dd>` (letters, digits, `.`, `_`, `:` and `-` only; write any other character, such as a space or the `/` in a branch name, as `-`), with a short word added if a second chat uses the same branch on the same day.
- After a context reset, call `list_agent_sessions` with that `product_id` and `mine: true`, and match the `locator`; if one matches, keep reporting to that session's `id` instead of starting a new one.
- Call `report_agent_session` at each milestone and whenever the plan or your next step changes, with the `plan`, `current_step`, `next_step` and any `decisions`; report at least every 15 minutes while working.
- When a person must decide, call `open_agent_session_question` with what is needed, the situation, the options with their consequences, your recommendation and what it blocks. Then call `wait_for_agent_session_answer` with `max_wait_seconds: 20`, and call it again each time it returns without an answer. While waiting, do nothing else. After 30 minutes without an answer, end the turn saying the question is still on the board, and call `wait_for_agent_session_answer` again the next time you run. When an answer arrives, act on it, then call `resolve_agent_session_question` with the decision.
- When you finish, call `end_agent_session` with the outcome (completed, failed or stopped), a short summary and links to what you produced.
- Never put secrets, credentials or your reasoning in a report. A session never completes a task: update the task itself as you normally would.
The priorify-agent-sessions skill teaches the same lifecycle in full.
Install
Release and marketplace packages
Platform-specific archives are published on the
latest release:
Clients that support Git or local ZIP installation can consume these packages
directly. ChatGPT availability requires approval and publication through
OpenAI's plugin directory; before that listing exists, the OpenAI archive is a
Codex/local-testing and marketplace-submission artifact. Other marketplace
listings will likewise be linked here only after approval.
Install from source
Clone the repository:
git clone https://github.com/toba-solutions/priorify-agent-kit.git
cd priorify-agent-kit
Then use the native workflow for your client:
Claude Code
claude --plugin-dir /absolute/path/to/priorify-agent-kit
Gemini CLI
gemini extensions install https://github.com/toba-solutions/priorify-agent-kit
Grok CLI
Keep the clone in a persistent directory, then validate and install it:
grok plugin validate /absolute/path/to/priorify-agent-kit
grok plugin install /absolute/path/to/priorify-agent-kit --trust
Review the plugin before accepting --trust. Start a new Grok session after
installation so its advertised skill list and MCP configuration are refreshed.
After Priorify.app publishes new MCP tools, run /mcps reconnect and start another
new session. Disabling and re-enabling the CLI entry leaves the current
session's cached tool list in place.
If grok inspect shows a pre-existing priorify-work skill, disable that legacy
skill at its source before testing this kit; it is not part of Priorify.app Agent
Kit and can otherwise compete with the four canonical skills.
Agent Skills fallback
Codex, Cursor, GitHub Copilot, Gemini CLI, and other compatible clients can
discover skills placed in .agents/skills:
mkdir -p .agents/skills
cp -R /absolute/path/to/priorify-agent-kit/skills/* .agents/skills/
For native plugin installers, point the client at this repository or use its
platform-specific ZIP from a release. See compatibility and installation
notes for the supported package layouts.
Connect Priorify.app
Every package points to the same Streamable HTTP MCP endpoint:
https://priorify.app/mcp
Installation and authorization are separate. On first use, your client may ask
you to sign in to Priorify.app and approve a connection. That OAuth connection does
not automatically grant access to every workspace or capability.
If an authenticated call returns connection_inactive, open
Priorify.app → Settings → Agent connections. Repeating OAuth will not repair an
inactive workspace binding.
Safety model
- Skills may recommend actions, but they do not authorize writes.
- Priority, status, assignment, budget, publication, and permission changes
require explicit user intent.
- Priorify.app's previews, expected versions, idempotency keys, capability grants,
and approval controls remain authoritative.
- Social planning records plans and lifecycle state; it does not publish to
social networks.
- Agent-work claim tokens are secrets and must never appear in progress text,
logs, or artifacts.
- No credentials or customer data are stored in this repository.
Repository layout
.
├── assets/ # Repository-owned visual assets
├── skills/ # Canonical portable skills
├── .codex-plugin/ # ChatGPT and Codex manifest
├── .claude-plugin/ # Claude plugin manifest
├── plugin.json # Agent Plugins v1 manifest
├── mcp.json # Agent Plugins MCP connection
├── .mcp.json # Native OpenAI and Claude MCP connection
├── gemini-extension.json # Gemini CLI extension manifest
├── scripts/ # Build and validation tooling
└── dist/packages/ # Generated release packages (ignored by Git)
The root skills/ directory is the only skill source of truth. Release bundles
are generated from it and must never be edited by hand.
Development
Requirements: Node.js 22 or newer. No package installation is required.
npm run check
This builds the four release layouts and verifies manifest versions, canonical
skill metadata, MCP endpoints, repository URLs, package contents, and the
absence of credential-like material.
Optional native validation:
python3 ~/.codex/skills/.system/plugin-creator/scripts/validate_plugin.py .
claude plugin validate .
gemini extensions validate .
See CONTRIBUTING.md before proposing changes and
docs/releasing.md before creating a version tag.
For independent client acceptance, give a tester the copy-paste brief in
docs/acceptance-test-prompt.md. It is designed
for Claude, Grok, and other tool-using agents and produces a comparable,
secret-free evidence report.
Support and security
License
The Agent Skills and packaging files in this repository are available under the
MIT License. The Priorify.app service, trademarks, and hosted product are
not granted under this repository license.
