FeatureBoard — the board Claude actually runs

A local, markdown-backed project board that lives inside Claude — packaged as a Claude Desktop / Cowork extension (.mcpb). Point it at a folder, and Claude plans, logs, prioritizes, and works tickets in plain language — then renders the whole thing as a live, interactive board artifact with a full analytics dashboard. No SaaS, no login, no lock-in. Your boards are plain featurelist.md / buglist.md files you own forever.

It started as a hosted web app (OpenClaw). We rebuilt it as an MCP server by taking the original app's file list and porting it, tool by tool, while learning MCP from scratch — and it came together fast. What you get is ~216 tools spanning the entire product surface: board and churn loop, analytics, testing, a media studio, CRM, leads, contracts, mail and marketing, a website builder with git deploys, and per-ticket commits. The board is the front door; everything the original app did is one sentence away in chat.

Why it's good

• It's the board Claude drives, not a database you poke. Ask "what's on my plate," "break FBF-300 into subtasks," or "ship the next ticket" — Claude pulls work, sets status, writes the code, logs the diff, and commits. The board stays honest because the agent keeps it honest.
• One command opens the real UI. get_board returns the shipped board as a Cowork artifact — Todo / In Progress / Done columns, filters, dark mode, click-to-move, and a 📊 analytics dashboard (velocity, timeline, bug health, work-log feed) — all wired live to the server. No hand-rolled mockups.
• Your files, your rules. Boards are byte-compatible markdown. Existing FeatureBoard folders work with zero migration. Writes are atomic. Delete the extension tomorrow and your boards are still just markdown.
• Depth when you want it. The same server that tracks tickets also generates media, runs a CRM, drafts contracts and campaigns, builds and deploys a marketing site, and commits your code per ticket. Turn it down to the essentials with one toggle, or open the whole toolbox.

Measured: what the board actually buys you

A paired agent trial (n=10, sonnet model tier) showed the work packet cuts diff churn 1.76x (837 vs 1,471 additions) with equal correctness (10/10 both). Token spend is 6–11% lower (77.5k vs 83.2k median), secondary to the focus gain: less to review, maintain, and scope drift on open-ended asks. See docs/EVAL-BOARD-VS-CHAT.md.

Quickstart

1. Install the extension: double-click featureboard-0.9.2.mcpb, or Claude Desktop → Settings → Extensions → Advanced → Install Extension.
2. On first run it asks for your Boards folder. Each subfolder with a featurelist.md or buglist.md is a board.
3. Say "open the FeatureBoard board" — the live board opens as an artifact. Then just talk:
   • "Log a bug: gallery modal won't dismiss on mouse-out."
   • "Brainstorm five onboarding features and add them."
   • "Move FBF-207 to done — fixed the parser regex."
   • "How's velocity looking this week?"

First-run tip: the extension defaults to Essential tools only (~34 core board tools) for a clean start. Turn that off in the extension settings to expose the full ~216-tool surface (CRM, media, website, marketing, and more). After installing or updating, fully restart Claude Desktop so new tools load.

Not using Claude? Cursor, Grok Build, and any other MCP client can run the same server over stdio — see docs/INSTALL-OTHER-IDES.md for per-IDE config, tool-count limits, and the IDE-neutral instructions flag.

Recipes

Automate your board's daily routines without a daemon. Copy-paste these prompts into Claude Code → Scheduled tasks to run nightly board health checks, weekly sprint close-outs, and daily standups. See docs/RECIPES.md for three ready-to-use scheduled task recipes and tool reference.

Everything it can do

Board & tickets — list_projects · create_project · list_tasks · get_task · add_feature · log_bug · add_features_bulk · import_tasks · validate_feedback · plan_work · update_task · set_status · decompose_feature · link_tasks · add_attachment · remove_attachment · delete_task · add_product · remove_product · get_project_config · set_project_config · get_ticket_history

The live board — get_board (returns artifact/board.html, ready to render as a Cowork artifact — this is what "open/show the board" calls)

Churn loop & orchestration — next_task · get_work_packet · log_work · get_agent_monitor · get_scratchpad · set_scratchpad · append_scratchpad (pull one ticket, assemble a focused brief, dispatch to a sub-agent, record the diff, repeat. get_agent_monitor is the live view of what's In Progress right now: elapsed time since each ticket started, its last event with age, token spend vs its cap:<tokens> label, and a stalled flag — an inactivity threshold, 30min by default. Stalled tickets also surface as a warning banner on the board, useful for keeping an eye on an unattended churn run. It also reports costSoFar/ capCost in dollars — see cost tracking below.)

Knowledge base — add_kb_doc · list_kb_docs · get_kb_doc · search_kb (a per-project kb/ folder of markdown docs, beyond the scratchpad — keyword-matched into get_work_packet's kbMatches automatically). rag_search is the retrieval layer over kb + repo docs + Done-ticket summaries: BM25 always, and HYBRID (BM25 + local MiniLM embeddings via the optional @xenova/transformers dep, reciprocal-rank fusion) when installed — the ~25MB model downloads once on first use then runs offline forever; without it, pure lexical, zero network.

Audit timeline — get_ticket_history (per-ticket chronological log: set_status/ update_task/assign_sprint append status, priority, label, sprint, and due-date change events to ticket_events.jsonl; get_ticket_history merges those with the ticket's work-log entries into one view. Tickets from before this existed just show their work-log history. The board's 🕘 button on each card renders the same timeline.)

Code review — get_ticket_diff captures the code a ticket produced (finds commits mentioning the ticket id in the project's code repo and returns per-commit summaries plus size-capped unified diffs from git show). add_review_comment leaves PR-style feedback on a ticket (optionally anchored to a file/line); list_review_comments and resolve_review_comment manage it. Unresolved comments surface in the ticket's next work packet (get_work_packet.reviewComments) so the next agent acts on the feedback, and a comment on a ticket in Review sends it back into next_task's queue — otherwise tickets awaiting review are held out of next_task (they belong to the reviewer). The board's 🕘 ticket panel lists review comments with their resolved state.

Analytics & health — get_metrics · get_health · get_work_log · predict_due_dates (get_metrics's velocity section rolls up $ cost by model too, alongside tokens. Cost tracking is model-aware: server/pricing.js ships defaults sourced from current Anthropic API pricing — input/output $/MTok per model tier (fable/opus/sonnet/haiku) — and every rate is overridable per project via set_project_config's pricing key, so a stale default is harmless. Work-log entries with an inputTokens/outputTokens split are priced exactly; older entries with only a tokens total fall back to a blended rate. get_agent_monitor and eval_report roll up the same $ figures per ticket/arm, and the board's metrics panel shows a cost line with a per-model breakdown.)

Timeline (piano roll) — get_timeline_data (FBMCPF-158). The board's 🎹 Timeline panel is an Ableton-style piano roll: every ticket is a clip on a horizontal lane, spanning its worked window (created → In Progress → completed, from ticket_events.jsonl with a completion-day fallback). Lanes group by product / model / type / sprint / status; a sticky time ruler adapts its granularity (hour / day / week / sprint); wheel zooms centred on the pointer, shift+wheel or grab-drag pans. A single datastream overlay (tokens / cost / additions+deletions) draws as a strip under the ruler. Hover a clip for a compact card (title, status, model, tokens); click to pin an expanded card that lazily pulls get_ticket_history for selectable cost / status-history / work-log fields.

Sprint close-out reports — close_sprint · get_sprint_report (FBMCPF-156) (close a sprint and turn its tickets + work log + metrics into four audience-specific reports — marketing (features shipped, positioning-ready copy), sales (customer-facing capabilities + the CRM tickets they resolve), technical (per-ticket changes, commits, ADRs touched), and executive (velocity, spend vs budget, health, risks). close_sprint refuses to close while a sprint still has open tickets unless force:true (open ones are then reported as carryover), writes deterministic reports/<sprint>/<audience>.md pads under the project (byte-stable, no model calls), and also returns a per-audience LLM prompt — the report packet as JSON plus an audience brief — so an agent can draft richer copy while staying grounded in real data. When the project has Slack configured it posts a one-block summary on close (a Slack failure is reported, never fails the close). get_sprint_report reads them back: no sprint → list sprints with reports; a sprint → its manifest and which audiences exist; sprint + audience → that report's markdown. The board's 📄 Reports panel lists every sprint with reports and renders each audience in its own tab.)

Board hygiene — scan_board_cleanup · prune_board (also lints open tickets missing a model:/cap: label — see intake orchestration guard below)

Intake orchestration guard (FBMCPF-159) — every ticket-creating tool (add_feature, add_features_bulk, log_bug, plan_work, import_tasks, validate_feedback in apply mode, and report_company_bug) fills in a model:/cap: label at intake whenever the creator didn't set one, via server/orchestration.js's deterministic type/keyword/product heuristics (bug vs. feature, docs/copy/rename → haiku or sonnet, architecture/UI-heavy/multi-file/parity → opus, conservative default sonnet + cap:80000). It never overrides a model:/cap: label you set yourself — only fills in what's missing. scan_board_cleanup then lints for any open ticket that still slipped through without one.

Code awareness — list_code_files · read_code_file · code_file_map

Testing & QA — suggest_test_stub · generate_test · generate_multi_model_tests · save_generated_test · list_test_variants · eval_model_matrix · bug_impact_scan · log_test_run · get_test_runs · test_runs_by_suite · get_regressions · save_test_page · list_test_pages · get_test_page · remove_test_page (multi-model generation fans one test prompt across model tiers — fable/opus/sonnet — then dedupes overlapping assertions so each bug gets one variant per model in the suite; eval_model_matrix is the empirical follow-up (FBMCPF-148) — it runs each tier's variant file against seeded, deterministic mutations of the target source in a temp dir (never the repo) and reports per-model defect-catch rate, unique catches, an overlap matrix, and cost per caught defect, to inform when downgrading test generation to a cheaper tier is safe)

Drift (ticket-vs-code fidelity) — drift_start · drift_record · drift_report · drift_remediate (score how well the code matches the ticket, report, and auto-remediate)

Media studio — save_media · list_media · get_media · edit_media · list_variations · revert_media · tag_media · annotate_media · remove_annotation · search_media · add_media_comment · list_media_comments · remove_media_comment · upload_reference · list_references · publish_media_to_site · draft_share · list_shares · remove_share

CRM & customers — add_company · list_companies · get_company · add_contact · add_crm_message · submit_crm_intake · list_crm_inbox · review_crm_message · customer_portal · link_customer_ticket · unlink_customer_ticket · ticket_customers · add_company_agreement · update_company_agreement · remove_company_agreement

Leads & field sales — add_lead · list_leads · set_lead_status · enrich_lead · convert_lead · leads_map · add_lead_area · list_lead_areas · add_lead_interaction · update_lead_location

Contracts — list_contract_templates · generate_contract

Mail & marketing — draft_email · list_mail · get_email · mark_email_sent · create_campaign · list_campaigns · get_campaign · record_campaign_open

Bookings — book_meeting · cancel_booking · list_bookings

Website builder + deploy — get_site · set_site · scaffold_site · edit_site_section · add_page · list_pages · remove_page · deploy_site · upload_site_asset · list_site_assets · set_site_analytics · set_analytics_config · auto_configure_analytics · get_site_traffic · enable_login_gate · disable_login_gate

Per-ticket git — get_git_config · set_git_config · commit_feature

Packaging — suggest_packaging · save_packaging_config · validate_packaging

Licensing — license_status · set_usage_type · activate_license · request_commercial_license · register_email

list_tasks is paginated and compact by default (limit, offset, compact:false) so large boards never blow the context budget. Write tools echo the exact post-write line. Tickets can carry an external ref (e.g. a plan item WI-1.2). See DESIGN.md for what was ported from the original app and what was intentionally left out.

The live board (Cowork artifact)

artifact/board.html is a self-contained Kanban that renders any board through the tools above — columns, cards with due dates / products / labels / refs, click-to-move status, a metrics strip, dark/light theme, and the 📊 analytics dashboard. get_board hands it to Claude ready for create_artifact, so any natural-language "open the board" surfaces the real UI (not a reinvented one) and refreshes from the server every time it opens.

Per-ticket git integration (optional)

FeatureBoard can commit — and optionally push — your code repo per finished ticket, exactly like the original app did. It's opt-in, off by default, and stores no secrets: pushing uses your machine's own git credentials (credential manager / SSH), just like running git yourself.

set_git_config     project=MyApp enabled=true branch=main push=true
set_project_config project=MyApp codeLocation="/path/to/repo"
commit_feature     project=MyApp ticket=FBF-42 title="Login flow"
#   → git add . && git commit -m "FBF-42: Login flow" && git push origin main

get_git_config shows current settings; commit_feature no-ops with a clear reason when disabled. Config lives in the board's git.config.json.

Pad mirror on close-out (graduated projects). Once a project's stage is graduated (see graduate_project), commit_feature and set_status ... Done also refresh a read-only snapshot of the pad — featurelist.md, buglist.md, scratchpad.md, agent_work_log.md, and config — into <codeRepo>/.featureboard/. commit_feature stages and includes the snapshot in the same commit as the code change; set_status Done refreshes it on disk even if commit_feature isn't called for that ticket. The central pad stays authoritative — this is a one-way snapshot, so cloning the code repo gets you the code plus the board history that shipped with it. Missing pad files are skipped; a mirror failure is reported as a warning and never blocks the commit or the status change.

Parallel dispatch in isolated git worktrees (optional)

For Cline-Kanban-style parallel agents, FeatureBoard can give each ticket its own git worktree — a separate checked-out directory on its own branch (ticket/<ticket>) that shares the repo's .git object store. N sub-agents can then work N tickets at once, each editing its own directory instead of fighting over one working tree, and you merge the branches back serially.

create_worktree  project=MyApp ticket=FBF-42          # -> <codeLocation>-worktrees/FBF-42 on branch ticket/FBF-42
list_worktrees   project=MyApp
cleanup_worktree project=MyApp ticket=FBF-42          # after the branch is merged back (refuses if dirty; force to override)

When a worktree exists for a ticket, get_work_packet includes a worktree block (path, branch, and step-by-step merge-back guidance) so the dispatched sub-agent edits there, never the shared repo. Board writes stay orchestrator-only, and branches merge back serially (checkout the base branch, merge/rebase ticket/<id>, run tests, commit_feature, then cleanup_worktree).

⚠️ Sync caveat — worktrees live OUTSIDE the code repo. Under Cowork, the host↔sandbox folder sync interacts badly with git's internal worktree administration files: a worktree created inside a synced repo mount can corrupt git internals or fail to sync. So worktrees default to a sibling directory <codeLocation>-worktrees/ (outside the repo), configurable with the project config key worktreeDir. create_worktree refuses a worktreeDir inside the repo and never auto-creates worktrees in the repo itself. Point worktreeDir at a path outside every synced mount (e.g. an OS tmpdir) if in doubt.

5-minute quickstart

1. Install: Claude Desktop -> Settings -> Extensions -> Install from file -> featureboard-mcp.mcpb. Restart Claude Desktop.
2. Approve tools once: Settings -> Extensions -> FeatureBoard -> Tool permissions -> Allow all (see next section).
3. Onboard: say "show my featureboard" — pick a license tier on the onboarding screen.
4. Plan: paste: "Plan work for project MyApp: create it and add tickets for " — plan_work files everything with model/cap labels.
5. Churn: say "work the board" — Claude pulls next_task, implements, commits, and closes tickets one at a time.
6. Review: "show the board" for the live Kanban; "close the sprint" for an audience-ready report.

That's the whole loop. Everything below is depth.

One-time setup: approve tools in bulk

FeatureBoard registers ~195 tools. In Claude Desktop, open Settings → Extensions → FeatureBoard → Tool permissions and choose Allow all for read and write once, instead of approving each tool the first time it fires — per-tool prompts stall agent churn loops mid-run. The onboarding screen reminds you of this on first use.

Licensing

FeatureBoard is source-available (see LICENSE.md): free for private non-commercial and public/nonprofit use (PolyForm Noncommercial), with a free 24-hour commercial trial after which write tools freeze (reads keep working) until a commercial key is activated. Commercial keys are US$99.99/seat/year (or US$9.99/seat/month), self-serve at https://featureboard.ai/buy. Activate with activate_license in either mode: paste the key itself, or pass the email + order id from your Polar purchase receipt and the server fetches the signed key for you from the featureboard.ai claim API — no copy-pasting a key required. You can also retrieve a key manually anytime with that same email + order id at https://featureboard.ai/claim and paste it in yourself. The board UI's "Activate license key" screen offers both paths side by side. On first use, onboarding asks which tier applies. Owners issue keys with the offline generator in owner/ (never shipped — .mcpbignore excludes it); see owner/README.md for the request → contract → issue-key → customer flow.

The same onboarding screen has an optional email field. Nothing is stored or sent unless you explicitly click "Save email" — that click is the only consent signal used. On submit, the email is saved locally and POSTed once to the featureboard.ai registrations listener; skip the field and it never leaves your machine.

FeatureBoard also collects anonymous usage telemetry: per-day counts of which of its own tools were called (names only — never arguments, results, or board content), batched to featureboard.ai at most once per 24h under a random install id that is never joined to your email. Opt out anytime with FEATUREBOARD_TELEMETRY=0 in the environment or set_global_config telemetry:false; verify with get_health. See docs/compliance/PRIVACY.md for the full disclosure.

Skills

FeatureBoard ships a Claude Skill in skills/ for the churn workflow this README describes:

• skills/featureboarding.skill — boards substantive dev requests automatically (no need to say "put it on the board") and churns the queue with parallel sub-agents per ticket dispatch blocks.

To install: drag the .skill file into Claude -> Settings -> Capabilities, or just open the file.

Data & storage

• Boards are plain markdown; a .featureboard/index.json sidecar caches the ticket counter. Markdown is always authoritative.
• Ticket ids are inferred from existing tickets in each file, so a board keeps its prefix regardless of folder name.
• Writes are atomic (temp-file + rename).

Build & develop

Requires Node 18+.

npm install                       # MCP SDK + zod
npm run check                     # syntax-check every server module
node --test                       # unit tests (1232)
npm run smoke                     # end-to-end stdio smoke test (no Claude Desktop needed)
npm run build && npm run bundle   # preflight, then pack featureboard-<version>.mcpb

Run the server directly over stdio:

FEATUREBOARD_DATA_DIR=/path/to/boards npm start

Nightly tests

npm run nightly runs the suite headlessly per nightly_tests.json and records a timestamped result under .featureboard/nightly/ (kept to keepRuns). The exit code mirrors the run so a scheduler can alert on regressions; if testLogPath points at a board's test_runs.md, the result is appended there so get_test_runs and the board's 🧪 Tests panel surface nightly runs next to manual ones. MCP servers have no daemon, so scheduling lives outside the server — point cron or Windows Task Scheduler at the script.

Troubleshooting

bash fails, the sandbox won't start, or a scheduled task silently stopped? That's almost always the Claude Desktop Cowork sandbox filling its disk — not a FeatureBoard bug — and it surfaces as useradd failed: exit status 12: cannot create directory /sessions/…, Workspace unavailable, or checks that never return. The fix is to reset the sandbox's disk image from your computer (your boards and repos are never touched). Full symptoms, the root cause, and step-by-step macOS/Windows reset instructions are in docs/TROUBLESHOOTING.md.

License

MIT © Lewis Valentine