⚡ sprinter

A self-hosted, gated work-backlog tracker with priority scoring, a filterable/sortable UI, and a built-in MCP server — so AI agents can pull the next item, do the work, and report progress back.

sprinter was built to drive an autonomous remediation loop: a fleet of agents queries the backlog over MCP (next_item), implements a fix, and marks the item done (update_item) — every change audit-logged. It works just as well as a plain human backlog board.

Why it exists

Most issue trackers are built for humans clicking around. sprinter is built so an agent can:

loop:
  item = mcp.next_item(category="security")   # highest-priority open item
  patch = implement_fix(item)
  mcp.update_item(item.id, status="fixed", notes="patch at ...")

…while a human watches the same data in a clean web UI (filter, sort, board, drill-down) and the database stays the single source of truth.

Features

• Priority scoring — every item carries Urgency / Impact / Effort (1–10); priority = U×I/E, so cheap-but-valuable work floats to the top. Re-scoring recomputes automatically.

• Filter & sort everything — by app, category, severity, status, free-text search, priority/effort ranges; server-side pagination (LIMIT/OFFSET + count) keeps it fast on thousands of items.

• Board view — kanban by status (open → in_progress → fixed → done).

• Apps registry — track items against the apps/repos they belong to (URL, repo path, stack, LOC, open/critical/high counts).

• Systemic themes — roll up issues that recur across many apps into portfolio-level initiatives.

• MCP server — 9 tools (list_apps, query_items, get_item, next_item, update_item, add_item, list_themes, stats, get_app) over JSON-RPC; bearer-token gated. Point any MCP client at it.

• Gated by default — the whole app sits behind a login (scrypt + HMAC-signed session cookie, login rate-limiting); the MCP endpoint requires a bearer token. Nothing renders pre-auth.

• Change log — every status/score/notes mutation is recorded with actor + timestamp.

Stack

FastAPI · PostgreSQL (any PG-wire-compatible engine works) · vanilla-JS single-page UI (no build step) · a dependency-light MCP JSON-RPC server. One small container + a database.

Quickstart

git clone https://github.com/danimoya/sprinter && cd sprinter
cp .env.example .env
# set DB_PASSWORD, SPRINTER_PASSWORD, SPRINTER_SESSION_SECRET (openssl rand -hex 32),
# and SPRINTER_MCP_TOKEN if you want the MCP endpoint enabled
docker compose up -d --build
docker exec sprinter python ingest.py     # loads the synthetic demo seed from ./demo
open http://localhost:8000                 # log in with SPRINTER_USER / SPRINTER_PASSWORD

The repo ships a synthetic demo dataset (demo/) — fictional apps and findings — so it's useful the moment it's up. Replace it with your own apps.json / backlog.json / themes.json (same shape) and re-run ingest.py.

MCP

Enable by setting SPRINTER_MCP_TOKEN, then register the endpoint with any MCP client:

claude mcp add --transport http sprinter http://localhost:8000/mcp \
  --header "Authorization: Bearer $SPRINTER_MCP_TOKEN"

next_item returns the single highest-priority open item — ideal as the head of an autonomous fix loop. update_item moves it through in_progress → fixed → done. All write tools require the token; reads do too.

Data model

• apps — key, name, urls[], repo_dir, profile, stack[], datastores[], loc, open_items, critical, high, max_priority

• items — id (stable hash), app, category (security|dast|infra|stub|improvement|feature), title, description, severity, cwe, owasp, location, evidence, remediation, confidence, urgency, impact, effort, priority, status, assignee, notes, …

• themes — cross-cutting initiatives spanning ≥3 apps

• change_log — audit trail of every mutation

Status lifecycle: open → in_progress → fixed (work done, pending deploy) → done (live), plus wontfix and rejected.

Security notes

• Set strong values for SPRINTER_PASSWORD, SPRINTER_SESSION_SECRET, and SPRINTER_MCP_TOKEN. With no password set, login is disabled (fail-closed). With no MCP token, the MCP endpoint returns 401.

• Put it behind a TLS-terminating reverse proxy in production and drop the ports: mapping (proxy by container name).

• The schema deliberately avoids UNIQUE-on-TEXT; idempotency is handled in ingest.py (full refresh) and update_item (by id).

License

MIT © Daniel Moya — see LICENSE.