Codex plugin, Claude Code plugin, and Model Context Protocol server

Find a shared root cause across different findings.

The board is the memory. The agent receives relevant systemic evidence before a fix and returns the verified outcome after the fix.

Use Engineering Board when an engineering agent must learn across sessions and domains. The product does not require a database, service, or daemon.

Install — free & MIT See a live board it runs

Three synthetic findings across hooks, board-view, and MCP connect into deterministic cluster C001 and a separate proposed root-cause hypothesis with cited evidence, an alternative explanation, and a falsifier. — Real output from the contained `/board-demo` sample · correlation remains separate from confirmation

Sanitized automated fixture showing 15 passed checks for deterministic cluster ranking, evidence-cited H001 lifecycle, rejected-claim memory, split and merge lineage, stale-plan refusal, offline parity, and escaped read-only rendering. — Real deterministic fixture output · 15 lifecycle checks · 2026-07-28

Sanitized automated fixture showing 11 passed checks for deterministic context retrieval, explicit fix outcomes, outcome-aware Learning confidence, value evidence, and rebuild equivalence. — Real deterministic fixture output · context before a fix · evidence after a fix · 2026-07-28

The project uses its public board. The project has CI tests, no runtime package dependency, and an MIT license. The server is on the MCP Registry.

Product differences

A single chat can hide recurrence. Engineering Board retrieves the common signal before an agent chooses another local correction. A cluster does not prove causation.

Systemic context before and after a fix

Task, path, and entry signals retrieve relevant clusters, hypotheses, negative memory, and Learnings. Each result includes a bounded canonical summary, its epistemic state, why it surfaced, and source references. PM summaries also surface medium/high-confidence Learnings matched to entries promoted in that pass.

Context → pattern evidence → root-cause choice → matched Learning

Memory that learns from outcomes

An explicit fix result updates H### history. A separate preview can revise L### outcome state and confidence. A failed fix can weaken or contest earlier memory.

Fix outcome → H### history → L### confidence

Claims for parallel agents

An atomic claim prevents two Worker agents from changing the same entry at the same time.

board-claim-acquire.sh

Falsifiable verification

Any suitable workflow can test a selected explanation. The optional TDD, review, and validation loop is one source of outcome evidence.

held · failed · partial · inconclusive

Stable pattern memory

A canonical P### identifier keeps a cluster stable after a label change. Preview and apply operations keep each correction explicit.

Markdown evidence, P### identity, and graph provenance

Use Claude Code Tasks and Engineering Board

Native Tasks store personal task state in ~/.claude/tasks/. Engineering Board stores shared and reviewable project memory in the repository.

Product comparison

This comparison uses observations from 2026-07-10. Each competitor name links to its source repository.

Tool	State is PR-reviewable markdown in your repo	Durable memory	Evidence-linked cross-domain pattern intelligence	Atomic claim-locking	Passive per-turn capture	Published team-visible board
engineering-board	Yes	Yes	Yes	Yes	Yes	Yes
beads · ~25k★	Partial	Yes	Partial	Yes	Partial	No
Backlog.md · ~6k★	Yes	No	No	Partial	No	Yes
Task Master · ~27.8k★	Partial	No	No	Partial	No	No
Claude Code native Tasks	No	Partial	No	No	No	Partial
claude-mem	No	Yes	Partial	No	Yes	No

beads has mature memory and claim features. Backlog.md has a detailed task model. Task Master converts a product requirements document into tasks. Engineering Board has a different primary function. It uses reviewable engineering evidence to show cross-domain patterns.

Screenshot of this repo's real rendered board: a search input, type/priority/status filter chips, and kanban columns — to do, review, validate, done — populated with entry cards. — the real thing — this repo's own board, as `/board-view` renders it · open it live

Install Engineering Board

Codex, Claude Code, and standalone MCP clients use one canonical Markdown board format. The project uses the MIT license.

Codex plugin

codex plugin marketplace add GhostlyGawd/engineering-board
codex plugin add engineering-board@engineering-board

# then start a new Codex session

The plugin supplies five skills and starts only the 19-tool Engineering Board MCP server. Pure repository-memory reads can run without a per-call prompt; write-capable tools remain approval-gated. It does not require another model-provider account.

Claude Code plugin

/plugin marketplace add GhostlyGawd/engineering-board
/plugin install engineering-board

# then, in a project:
/board-setup

# optional contained first win:
/board-demo

The plugin also registers the MCP server through .mcp.json.

MCP server for an MCP client

# one line from PyPI
claude mcp add engineering-board -- uvx engineering-board-mcp

# fallback — run from a clone:
git clone https://github.com/GhostlyGawd/engineering-board
claude mcp add engineering-board -- \
  python3 "$(pwd)/engineering-board/mcp-server/engineering_board_mcp.py"

# Claude Desktop: claude_desktop_config.json
"mcpServers": {
  "engineering-board": {
    "command": "uvx",
    "args": ["engineering-board-mcp"]
  }
}

The Python server has no package dependency. It provides 19 tools through standard input and output. Read the client setup procedures.

An uninstall operation does not remove the Markdown board. The system treats scratch content as untrusted data. Read the security policy.