SpecForge

Local-first requirements engineering and specification management for software and hardware projects.

SpecForge tracks every artifact of a project — ideas, requirements, decisions, tasks, tests, and verification evidence — as plain Markdown files in a git repository. It enforces traceability between them, runs a release gate to confirm every requirement is verified before you ship, and integrates with AI tools so Claude Code can read and write your project during development.

                   ┌─ exploration ─┐
ideas ──► candidates ──► requirements ──► tasks ──► verified ──► release
                              │                         │
                         decisions               verifications
                         assumptions
                         constraints

Why SpecForge?

Most teams track work in one tool (Linear, Jira, GitHub Issues) and requirements in another (Confluence, Google Docs, Notion) — or nowhere at all. Nothing connects. When something breaks in production, reconstructing why a decision was made, what requirement it served, and what test was supposed to catch the problem is an archaeological dig.

SpecForge solves this by treating every piece of project knowledge as a first-class artifact with a permanent ID, a status, and typed links to other artifacts. The result is a queryable, auditable, AI-readable project graph stored in plain files that any tool — your editor, your CI pipeline, Claude Code — can read.

See docs/THEORY.md for the requirements engineering principles behind the design.

Feature summary

Installation

# Clone and install in editable mode
git clone <repo> specforge
cd specforge
pip install -e .                     # core + CLI + daemon
pip install -e ".[desktop]"          # + PySide6 desktop app
pip install -e ".[dev]"              # + pytest, ruff, mypy, httpx

# Verify
specforge --help
specforge-daemon --help

Quick start

1. Create a project

specforge init ./my-project --git --name "My Project"

This creates the full directory scaffold, initialises git, and writes a commented .specforge.yaml configuration file.

2. Capture an idea

specforge add-idea ./my-project "Offline support" \
  "Users lose work when the network drops."

3. Promote to a requirement

specforge promote ./my-project IDEA-0001 requirement \
  --text "The application shall serve cached content when no network
  connection is available, with no data loss on reconnect."

4. Add a linked task

specforge add-task ./my-project "Build offline cache layer" \
  --text "Implement SQLite-backed cache for the last 50 items.
  Sync to server on reconnect with conflict detection." \
  --implements REQ-0001 --git

5. Mark implemented and verify

specforge update-status ./my-project TASK-0001 implemented
specforge add-test ./my-project "Offline read test" \
  --text "Disable network. Open app. Assert cache loads < 2s." \
  --req REQ-0001
specforge add-verification ./my-project "Offline cache — QA passed" \
  --text "QA on iOS 17 and Android 14. All scenarios passed." \
  --req REQ-0001 --test TEST-0001 --git
specforge update-status ./my-project REQ-0001 verified --git

6. Check the release gate

specforge check ./my-project    # PASS / FAIL
specforge report ./my-project   # acceptance report

AI drafting

Configure a provider once:

specforge config ./my-project --set llm.provider=anthropic
# reads ANTHROPIC_API_KEY env var automatically
# or: --set llm.provider=ollama  (no API key needed, runs locally)

Draft any artifact from a plain-language prompt:

specforge draft ./my-project requirement \
  "Passwords must be at least 12 characters and contain a mix of types" \
  --title "Password complexity policy"

SpecForge calls the LLM, displays the generated body in a Rich panel, and asks for confirmation before creating the artifact. If the remote provider fails, it offers to retry with local Ollama.

Web UI

specforge-daemon     # start on http://127.0.0.1:8765
# open http://127.0.0.1:8765/ui

The web UI supports dark, light, and system colour modes. Click ? Help for built-in documentation on every feature.

For development (auto-restart on source changes):

SPECFORGE_DEV=1 specforge-daemon

Claude Code / MCP integration

specforge mcp-config /path/to/project    # print config snippet

Paste the printed mcpServers block into Claude Code settings.json. Claude can then call all 15 SpecForge tools directly in conversation: create artifacts, search, check status, promote, verify.

Available MCP tools: create_artifact, promote_artifact, update_status, get_artifact, list_artifacts, link_artifact, unlink_artifact, search, get_status, context_pack, validate. Additional workflow tools: check_gate, draft_artifact, bulk_update, list_templates.

Architecture

specforge_core/         Core library (no HTTP, no UI)
  models.py               Artifact, ArtifactKind, ArtifactStatus
  project.py              Project filesystem operations, ID generation
  trace.py                SQLite trace index (full-text search, graph queries)
  validation.py           Link and status consistency checks
  search.py               Multi-term AND search
  export.py               CSV and Markdown traceability matrix
  report.py               Acceptance report generator
  contextpack.py          AI context pack builder
  status.py               Project health dashboard
  adapter.py              Tool-call dispatcher for AI agents
  config.py               .specforge.yaml loader/saver
  llm.py                  Multi-provider LLM client (Anthropic/OpenAI/Ollama)
  bulk.py                 Batch artifact operations
  templates.py            Artifact template loader
  webhooks.py             HTTP webhook dispatcher
  plugins.py              Local Python plugin loader

specforge_cli/          Command-line interface (Typer)
  main.py                 All CLI commands

specforge_daemon/       Local REST API (FastAPI + uvicorn)
  api.py                  All REST endpoints + help content serving
  main.py                 uvicorn entry point
  mcp_server.py           MCP stdio JSON-RPC server
  watcher.py              watchdog auto-rebuild on file changes

specforge_studio/       Desktop GUI (PySide6)
  main.py                 Main window, dialogs, help viewer

specforge_web/          Browser client
  index.html              Single-page app (no build step)
  help/                   13 Markdown help files

All state lives in the project directory. The daemon and studio are thin clients that delegate all logic to specforge_core.

Configuration

.specforge.yaml at the project root:

project_name: "My Project"
git_commit: false           # auto-commit every write

llm:
  provider: anthropic       # anthropic | openai | ollama
  model: claude-sonnet-4-6  # blank = provider default
  api_key: ""               # blank = ANTHROPIC_API_KEY env var
  base_url: ""              # blank = provider default

webhooks:
  - url: https://hooks.slack.com/services/XXX
    events: [artifact.created, artifact.promoted]
    secret: "optional-hmac-secret"

specforge config ./proj                          # show config
specforge config ./proj --set git_commit=true    # update key
specforge config ./proj --set llm.provider=ollama  # nested key

Documentation

Built-in help system: start the daemon, open the web UI, and click ? Help — or run specforge --help / specforge <command> --help.

Running tests

pytest                    # 336 tests
pytest tests/test_llm.py  # specific module
pytest -q                 # quiet mode

Full lifecycle demo

./scripts/create_demo_project.sh    # 22-step end-to-end demo

This runs a complete project lifecycle — init, ideas, requirements, decisions, tasks, tests, verification, export — and leaves the resulting project in examples/demo_project/ for inspection.