work

Project management: projects, epics, deliveries, tasks, business units, and member placement against the work schema.

Endpoint	https://work.mcp.devfellowship.com/mcp
Tools	34
Backing data	work schema (projects, epics, deliveries, tasks, business units, placements).

Connect

Add https://work.mcp.devfellowship.com/mcp to your client (see Getting Started) with your Authorization: Bearer <token> header.

Resolving members

lookup_member (name → member id) now lives on the learn endpoint. Use it there to get the member_id you pass to create_placement.

Projects

Tool	Description
create_project	Create a new project.
delete_project	Delete a project. This will fail if the project has associated epics.
get_project	Get a specific project by ID with its epics.
list_projects	List all projects with optional filters.
update_project	Update an existing project.

Epics

Tool	Description
create_epic	Create a new epic. ⚠️ An epic cannot be shown on a plan — the plans-app renders bound tasks only. Bind the epic’s tasks instead.
delete_epic	Delete an epic. This will fail if the epic has associated deliveries or tasks.
get_epic	Get a specific epic by ID with its deliveries and tasks.
list_epics	List all epics with optional filters.
update_epic	Update an existing epic.

Deliveries

Tool	Description
create_delivery	Create a new delivery.
delete_delivery	Delete a delivery.
get_delivery	Get a specific delivery by ID with its epic and tasks.
list_deliveries	List all deliveries with optional filters.
update_delivery	Update an existing delivery.

A task created for a plan MUST also be bound to it

Creating a task and putting it on a plan are two calls on two different MCPs. This server writes work.tasks. A plan renders its execution checklist from work.entity_connections, and nothing here writes that table.

work MCP   create_task(...)                          -> the task exists
plans MCP  set_plan_tasks({ slug, tasks: [{task_id}] }) -> the plan SHOWS it

set_plan_tasks replaces the plan’s task set — always send the full desired list, not only the new ids.

The failure is silent. A created task is a complete, valid row on its own, so skipping the second call raises nothing. The only symptom is an empty rail on the plan page, which reads as “no work has started” rather than “the link is missing”. That is exactly what happened on 2026-08-12: 18 tasks created under an epic, reported as linked, plan page empty.

Verify by reading the plan back and seeing your tasks in its links.

Tasks

Tool	Description
create_task	Create a new task. Requires a context package (why/what) and a stage_id (spec / design / execution / review / decision / qa) — a task without a stage is invisible on the board. If it belongs to a plan, also call set_plan_tasks on the plans MCP — see the callout above.
delete_task	Delete a task.
get_task	Get a specific task by ID with its epic and delivery.
list_tasks	List all tasks with optional filters.
update_task	Update an existing task.

Business Units

Tool	Description
create_business_unit	Create a new business unit. Requires finance/admin/owner role.
delete_business_unit	Delete a business unit by ID. Requires finance/admin/owner role.
get_business_unit	Get a specific business unit by ID. Requires finance/admin/owner role.
list_business_units	List all business units with optional filters. Requires finance/admin/owner role.
update_business_unit	Update an existing business unit. Requires finance/admin/owner role.

Placements

Tool	Description
create_placement	Record a DFL fellow placement (job/freelance/internal) in work.placements. Resolve the member_id first via lookup_member on the learn endpoint.
create_member_label_alias	Map a raw speaker/label string to a member by upserting into work.member_label_aliases.

Meetings

Tool	Description
merge_meetings	Dedup a duplicate recurring-meeting cluster by merging a skeleton meeting into a canonical meeting, then removing the skeleton. Lossless on duplicate-date collisions: it never drops a transcription-bearing occurrence in favor of an empty one. For a date both meetings share, it compares transcripts on both sides — if the canonical occurrence is empty but the skeleton one carries a real transcript, it re-parents the skeleton occurrence onto the canonical and deletes the empty canonical duplicate (the real transcript survives). If both carry transcripts it reports a conflicts entry, keeps both, and does not remove the skeleton row (skeleton_meeting_removed: false, manual review). Empty skeleton duplicates of an existing canonical date are dropped; skeleton occurrences on a new date are re-parented. Other meeting_id child rows (participants/concepts/expertise/tools/underlines) are re-parented onto the canonical, guarding the work.meeting_participants unique constraint. The skeleton row is removed only when there are zero unresolved conflicts. Transcriptions live per-occurrence so nothing is concatenated and the canonical survives. It also preserves the work.meetings.template field: if the canonical is null/"default" and the skeleton carries a non-default template (e.g. "weekly_leaders"), the canonical is updated to the skeleton’s template (carry the more-specific value forward; never downgrade). Pass dry_run: true first to preview the change without writing.
set_meeting_recurrence_time	Set the time-of-day window (recurrence_start_time / recurrence_end_time) on a single recurring meeting series in work.meetings, so its recurring occurrences render at the right hour on the /meetings time-grid. Times are 24h wall-clock "HH:MM" or "HH:MM:SS" (Postgres time). The meeting must already be a recurring series (is_recurring = true); recurrence_day_of_week is set separately. Returns the updated row.
set_meeting_template	Set the UI template (work.meetings.template) on a single meeting. The template selects which structured-underline UI the dfl-learn /meetings view renders — the weekly-leaders structured form is gated on template = "weekly_leaders". Allowed values: "default" | "weekly_leaders" (mirrors the dfl-learn UnderlineTemplate enum); unknown values are rejected. Returns the updated row. Primary use: restore a meeting’s template after a dedup-merge collapsed it to "default".

Meeting series data-ops

Data operations live here, not in dfl-schema

Tainan, 2026-07-28: “prs nao sao para remover rows ou inserir seeds. pra isso use mcp. crie se nao existir”. PRs in devfellowship/dfl-schema are for schema (DDL). Removing rows, seeding, and consolidating duplicates are data operations and go through these tools, on the caller’s user-JWT (RLS enforced, no service_role).

Transcripts live per-occurrence, and all 7 FKs onto work.meetings(id) are ON DELETE CASCADE — so deleting a series parent before its occurrences are dealt with silently destroys transcripts. An MCP tool cannot hold a Postgres transaction across calls, so order is the safety mechanism: promote → verify → delete. Always run dry_run: true first on prod.

work.meetings also has an AFTER INSERT trigger (meetings_to_n8n) that POSTs every new row to the n8n process_devfellowship_meetings webhook. A user-JWT tool cannot disable it, so each promoted meeting fires one n8n workflow — the count is reported as n8n_webhook_inserts in both dry-run and execute results.

Tool	Description
list_meeting_series	Read-only discovery of recurring series parents (work.meetings with is_recurring = true), so dedup targets can be found without psql. Filters: created_after / created_before (ISO-8601), title_contains, fireflies_calendar_id_contains, limit. Each result carries id, title, fireflies_calendar_id, meeting_date, created_at plus occurrence accounting (occurrence_count, occurrences_with_transcript, occurrences_with_transcript_id); include_occurrences: true adds the per-occurrence rows (id, date, status, has_transcript, fireflies_transcript_id). Transcript text is never returned — it is tens/hundreds of KB per occurrence. Sorted by created_at.
remove_meeting_series	Remove one recurring series parent, promoting its occurrences to standalone (“avulsa”) meetings first so the cascade can never eat a transcript. occurrence_policy: promote_real_discard_empty (default — promote occurrences carrying transcript text or a Fireflies transcript id, delete the empty materializer placeholders), promote_all, or require_empty (pure safety mode). Guards, all of which abort rather than guess: occurrences with transcript text but no fireflies_transcript_id (unverifiable); occurrences whose transcript id already exists on a meetings row (skipped, not re-inserted — the partial unique uq_meetings_fireflies_transcript_id, which also makes the tool idempotent); rows in meeting_participants / meeting_concepts / meeting_expertise / meeting_tools / meeting_transcript_segments / meeting_underlines referencing the parent or its occurrences (use merge_meetings, which knows how to re-parent them); and a post-promotion assert that every surviving occurrence has a standalone counterpart before anything is deleted. Promoted rows get is_recurring: false, fireflies_calendar_id: null (deliberately detached), the parent’s title, and a meeting_date rebuilt from the occurrence date plus the parent’s time-of-day. Refuses a non-recurring meeting unless allow_non_recurring: true. Returns counts (total_occurrences / promoted / discarded / already_present, which must balance), the created standalone ids, and n8n_webhook_inserts.
promote_occurrence_to_standalone	Promote one work.meeting_occurrences row to a standalone meeting and detach/delete the occurrence — the generic primitive underneath remove_meeting_series. Optional title_override; keep_occurrence: true creates the standalone but leaves the occurrence in place (copy-then-review, and the escape hatch for an occurrence whose transcript has no id or that still has participant/underline rows attached). Same unique/transcript/cascade guards; insert-and-confirm happens before any delete.

Weekly updates

Read-only. What one member did in the last N days, per source, with a daily text ready to paste. Mirrors the Weekly updates page of dfl-learn (/weekly-updates).

Tool

Description

get_weekly_updates

member_id? (default: you), days 1–14 (Brasília calendar days, today included). Sections: tasks moved to done / dev_completed (work.tasks.status_timestamps), tasks in progress, merged PRs (review_request.review_requests.merged_at, by created_by), Core Daily lines (resolved per meeting: segment attribution, then the top alias, then the exact name; any ambiguity leaves the line out), plans whose status changed (plans app, with your own session). Returns unavailable[] and truncated[] per source and daily_text in the Bom dia / Ontem / Hoje / No blockers format, one line per item, with no PR numbers, repos, task identifiers or links. Everything runs on your JWT; no service role.