tensorpm-agentic-pm

TensorPM Skill

Use this skill to operate TensorPM as the dedicated project management subagent — via MCP for structured CRUD and A2A for high-level, context-aware operations on a running TensorPM desktop app.

TensorPM has two interfaces over the same project graph:

• Desktop app for humans: classical project management UI plus a built-in "TensorPM" agent that takes PM busywork off the user (extracting action items from documents, proposing context updates, suggesting next steps).
• MCP/A2A for external agents: this skill. Claude Code, Codex, OpenClaw and any MCP/A2A client read and write the same project graph (goals, action items, decisions, history) — so they don't have to re-explain or hallucinate project context.

TensorPM itself is free. For AI capabilities outside MCP (A2A), use your own API key (BYOK) or create an account.

When To Use

• You need to list, create, or update TensorPM projects or action items.
• You need to switch/list workspaces.
• You need to set AI provider keys through TensorPM (set_api_key).
• You need conversational project-level changes via A2A (message/send).
• You need to make external MCP servers available inside TensorPM's AI chat.

When Not To Use

• The request is only about website/account/billing pages.

Installation (Agent CLI)

Preferred package-manager install:

# macOS
brew install --cask neo552/tensorpm/tensorpm

# Windows (PowerShell)
winget install --id Neo552.TensorPM --exact --accept-package-agreements --accept-source-agreements

Direct-download fallback (works on every platform, no installer script required):

# macOS DMG
curl -fL -o /tmp/TensorPM-macOS.dmg \
  https://github.com/Neo552/TensorPM-Releases/releases/latest/download/TensorPM-macOS.dmg
hdiutil attach /tmp/TensorPM-macOS.dmg -nobrowse -quiet
cp -R "/Volumes/$(ls /Volumes | grep -i TensorPM)/TensorPM.app" /Applications/
hdiutil detach "/Volumes/$(ls /Volumes | grep -i TensorPM)" -quiet

# Linux AppImage
curl -fL -o ~/TensorPM.AppImage \
  https://github.com/Neo552/TensorPM-Releases/releases/latest/download/TensorPM-Linux.AppImage
chmod +x ~/TensorPM.AppImage

# Windows installer
Invoke-WebRequest -Uri https://github.com/Neo552/TensorPM-Releases/releases/latest/download/TensorPM-Setup.exe -OutFile $env:TEMP\TensorPM-Setup.exe
Start-Process $env:TEMP\TensorPM-Setup.exe

If the agent needs to verify or pin a version, list releases via gh release list -R Neo552/TensorPM-Releases (the GitHub /releases/latest endpoint excludes pre-releases, so prefer gh release list when only beta tags exist).

Runtime Prerequisites

1. Start TensorPM desktop app.
2. For MCP usage with external AI clients: ensure client integration is installed once (via Settings -> Integrations or A2A POST /integrations/mcp/install).
3. For A2A usage: verify local endpoint http://localhost:37850.

External MCPs Inside TensorPM

When a user asks to give TensorPM access to external MCP servers, write an MCP config file and let TensorPM import it on startup.

Preferred paths:

1. User-wide: ~/.tensorpm/agent-mcps.json
2. Project-local: .tensorpm/agent-mcps.json
3. Local override: .tensorpm/agent-mcps.local.json
4. Explicit override: set TENSORPM_AGENT_MCP_CONFIG_FILE to one or more paths separated by the OS path delimiter.

TensorPM accepts standard mcpServers config blocks and TensorPM-native agentMcpServers. See Agent MCP Clients for schema, examples, and safety notes. The UI exports its own generated snapshot to ~/.tensorpm/agent-mcps.generated.json; agents should write user-managed config to ~/.tensorpm/agent-mcps.json.

MCP vs A2A Routing

Task	Use	Why
Structured action-item CRUD	MCP tools	Direct typed operations
Set provider API keys	MCP set_api_key	Dedicated secure write-only tool
Project-wide/contextual changes	A2A message/send	Managed by project manager agent
HTTP-based automation/client integration	A2A REST/JSON-RPC	Endpoint-first integration path
Multi-turn planning with conversation state	A2A with contextId	Native conversation continuity

Rule of thumb:

• Prefer MCP for explicit CRUD operations.
• Prefer A2A for high-level intent and context-aware planning.

Minimal Workflow

1. Verify TensorPM is running.
2. Choose MCP vs A2A via the routing table above.
3. Execute operation.
4. Read back result (list_*, get_project, or A2A read endpoint) to confirm state.
5. Summarize applied changes and any follow-up action.

References

• MCP Tools: tool catalog and usage boundaries.
• Agent MCP Clients: configure external MCP servers for TensorPM's AI chat.
• A2A API: discovery, JSON-RPC methods, REST endpoints, examples.
• Action Items & Dependencies: fields, dependency types, payload examples.

Notes

• IDs are UUIDs.
• Dates use ISO format (YYYY-MM-DD).
• propose_updates requires human approval before apply.
• MCP and A2A operate on the same local TensorPM data.
• Release notes: https://github.com/Neo552/TensorPM-Releases/releases/latest