Assistant Knowledge base Views Automation Teams Deploy ← PineCompute Request access
Now in pilot · Self-hosted · Local AI

Your projects. Your AI.
Your hardware.

PineProject is a project board with a built-in assistant that answers from your cards, comments and documents, cites every source, and says so when it doesn’t know. It runs on your machine, and your data never leaves it.

100% on-premise No cloud, no telemetry Runs on a laptop, no GPU required
Atlas v2 Launch · team board
KanbanCalendarGridGantt
live sync
In Progress 2
Beta invite emails
MarketingOct 11MR
Release notes draft
DocsOct 12JS
Due Tomorrow ⚡ +1d
Vendor security review
OpsOct 8KG
Due Today ⚡ today
Done 2
Load test checkout API
EngOct 6JS
Finalize pricing page
MarketingOct 4MR
⚡ AUTOMATION Moved “Vendor security review” to Due Today
5views of one board
6document formats indexed
0bytes sent to the cloud
1process, one port
430+automated tests
The problem

Project knowledge shouldn’t cost you your privacy.

The answer to “why did we decide that?” is somewhere in a card, a comment thread or the steering-committee minutes. The tools that can find it want you to upload all of it first.

01 / EXPOSURE

Cloud AI reads your roadmap

Every question you ask a hosted assistant sends your plans, vendors and client names to someone else’s servers, where they may be logged and retained.

02 / SCATTER

Decisions get lost

The task lives on the board, the reasoning lives in a .docx on the shared drive, and the context lives in someone’s head. Finding the “why” turns into an archaeology project.

03 / EXCLUSION

Air-gapped teams get nothing

Regulated, defense and security-conscious teams can’t use cloud AI at all, so they’re left searching hundreds of files by hand.

Knowledge Assistant

It cites, or it refuses.

The assistant only sees passages retrieved from your board and your documents. If nothing relevant clears the bar, it says so, and the language model is never even called.

01

Retrieve

Hybrid search runs keyword (BM25) and semantic vectors together and fuses the rankings. Every board is isolated, and your shared documents are visible to all boards.

02

Gate

A similarity floor, tuned against measured data, drops weak matches before the model sees anything.

03

Answer

The answer streams token by token. Sources are attached from the retrieval list itself, not parsed out of the model’s prose, so a citation can’t be invented.

∅

Or refuse, honestly

With no grounded evidence you get a plain “I couldn’t find that” instead of a confident guess.

Why did the beta slip to Oct 14?
● Answered · 2 sources
The load test on the checkout API found p95 latency over budget 1, and the 9/26 steering minutes approved a one-week slip so the fix could ship first 2.
[1]Load test checkout API card · comment
[2]steering_minutes_0926.docx kb file
What’s the budget for Q3 contractors?
○ Not found
I couldn’t find that in this board or the knowledge base. Try rephrasing, or add the document that covers it.
Brain-as-a-swap: Ollama · qwen2.5:7b any OpenAI-compatible endpoint
/mnt/shared/Atlas
  • docxproject_charter.docx✓ indexed
  • docxsteering_minutes_0926.docx• not indexed
  • xlsxrisk_register.xlsx✓ indexed
  • mdadr-012-queue-choice.md✓ indexed
  • pdfvendor_msa_signed.pdf✓ indexed
  • csvload_test_results.csv• not indexed
  • txtold_kickoff_notes.txt✓ indexed
Last scan: yesterdayPress Rescan →
KB-Pro

Point it at the shared drive.

Charters, risk registers, ADRs and meeting minutes already live in folders. An admin picks the folder, presses Rescan, and the assistant can answer from those files with the same cite-or-refuse guarantees.

Six formats: .md .txt .csv .docx .pdf .xlsx

Change detection by content hash. Unchanged files are skipped, edited ones re-indexed, and deleted ones purged from the index.

Fail-soft per file. One corrupt spreadsheet is reported and skipped, and the rest of the scan carries on.

Views

One board. Five ways to see it.

Cards carry start dates, durations, priorities, assignees and dependencies, so the same data drives a kanban wall, a calendar, a spreadsheet grid and a Gantt chart. Try it:

Automation & agents

The board keeps itself current.

Dates move cards for you, the assistant writes your standup, and your own AI agents can work the board through a standard protocol.

Date columns

Tag a column “Due Today”, “Due in 3 days” or “Overdue” and incomplete cards move in on their own. The move runs at the daily tick, or on the next board load if the server was off. Done and undated cards stay put.

activity · ⚡ Automation Vendor security review Due Tomorrow → Due Today

Standup digests

“What changed since yesterday?” is generated from the board’s activity log: moves, completions, new comments. If no model is running it falls back to a plain template, so you still get your digest.

## Atlas v2 · since yesterday - 1 completed: Load test checkout API - 2 moved, 3 new comments - Due today: Vendor security review

MCP server

Connect Claude, Cursor or any MCP-capable agent. Each tool is a thin, authenticated call into the same API under a user-scoped token, so agents get exactly that user’s permissions.

list_boardsget_boardcreate_cardmove_cardadd_commentsearch_knowledge
Individual → Team

Start solo. Bring the team when you’re ready.

One install, two modes. Switch by re-running the setup wizard.

Individual mode

Your laptop. Zero login.

A private workspace on localhost for one person. Open the browser and start planning.

  • Bound to this computer only, so no accounts are needed
  • All five views, the assistant and KB-Pro included
  • SQLite file database with zero ops
Team mode

Your LAN. Live together.

Share the board across the office network. Everyone sees moves the moment they happen.

  • Login on every route, including the live-sync socket
  • Live board sync over WebSockets: moves, edits and comments show up instantly
  • Comments, @mentions, assignees and a per-board activity feed
viewer‹member‹admin‹owner

Duplicate as template

Clone a board’s columns and labels, with or without its cards and dates.

SQLite → Postgres

Outgrow the file database with one migration script that verifies row counts.

Mutations never wait on AI

Indexing runs after commit, in the background. If the model is down, the board still works.

Deploy

Three commands. One port.

A setup wizard asks for the mode, network and port, generates secrets, creates the owner account, and checks for Ollama and Node before anything goes live. One process serves the API, the web app and live sync.

Runs on
Linux workstation or server
AI runtime
Ollama, local (optional)
Database
SQLite or Postgres
Network
Localhost, LAN, or fully air-gapped
~/PineProject
$ ./pp.sh setup
? Mode          Team
? Network       LAN
? Port          8796
? Owner email   pm@yourco.internal
✓ JWT secret generated   ✓ Ollama found
✓ Node 22                ✓ saved .pp/config.env

$ ./pp.sh build
✓ web UI built

$ ./pp.sh start
● PineProject is running
  local  http://127.0.0.1:8796
  LAN    http://10.0.4.20:8796
Pilot program

Put your projects on your hardware.

PineProject is in pilot with a small group of teams. Request access and we’ll set up a guided demo, then help you install it on your own machine.

Product & engineering teams

A shared board with an assistant that remembers why decisions were made.

Regulated & air-gapped orgs

Modern AI-assisted project management with nothing leaving the network.

Solo PMs & consultants

Individual mode on your laptop: private, fast and free of login screens.

Prefer email? [email protected]

We reply within 48 hours. No newsletter, no spam.