Calaf MCP Server
Connect Claude, ChatGPT, or any MCP-capable assistant to Calaf — the MBA recruiting manager. Your assistant researches target firms at whatever depth you ask, writes and validates structured "seeds," and files them into your recruiting book. People it proposes always arrive staged for your review — nothing enters your network behind your back.
Two servers, one product. ai.calaf/calaf connects to your account (OAuth). ai.calaf/seeds is the no-auth authoring loop — the format and the validator, usable by anyone, no Calaf account required.
| Server | Endpoint | Transport | Auth |
|---|
| Calaf | https://calaf.ai/api/mcp-account | Streamable HTTP | OAuth 2.1 (dynamic client registration) |
| Calaf Seed Tools | https://calaf.ai/api/mcp | Streamable HTTP | None |
Things to ask once connected
- "Research 30 lower-middle-market PE funds that hire MBA associates, validate the seed, and import it into my book."
- "Read Calaf's seed format, then turn the target list from this conversation into a seed and validate it until it's clean."
- "List my Calaf workspaces and import these interview prep questions into the consulting one."
- "I just got off a superday with Evercore — log it, move my application, and set a follow-up task for Friday."
- "Create a workspace for lower-middle-market PE, add my ten targets, and draft outreach templates."
- "Research the partners at my five priority funds and add them to my book" — they'll land as a batch waiting for your review in Calaf, never as silent rows.
Installation
claude.ai (web / desktop)
- Open Settings → Connectors (claude.ai/settings/connectors)
- Add custom connector → name it
Calaf, URL https://calaf.ai/api/mcp-account, leave Client ID/Secret empty
- Click Connect → you'll land on Calaf's consent screen — sign in and approve
Requires a paid claude.ai plan (custom connectors are Pro and above).
Claude Code
claude mcp add --transport http calaf https://calaf.ai/api/mcp-account
Any other MCP client
Add a remote (Streamable HTTP) server at https://calaf.ai/api/mcp-account. The OAuth flow is standard: dynamic client registration + PKCE; your client discovers the authorization server via RFC 9728 protected-resource metadata.
No MCP client at all
The copy-paste loop works with every assistant: Calaf's Guide and Settings carry a one-paste prompt that teaches any AI the seed format; the assistant can self-check via the open validator API (below) and you import the file at Settings → Workspace → Seeds.
Returns the complete calaf_seed: 1 specification (markdown) — sections, ceilings, parser behavior, and the rules an assistant should follow when authoring a seed. Read it before authoring; then check drafts with validate_seed.
No parameters.
validate_seed · read-only
Runs a draft seed document through the exact parser the Calaf app imports with and returns the verdict: per-section counts plus the precise issue list. Issue severity error means nothing would import; dropped means that row or field is discarded and the rest lands. Fix and re-validate until clean before delivering the file.
| Parameter | Type | Required | Description |
|---|
seed | string | object | yes | The complete seed JSON document — as text, or as the JSON object itself |
Returns (structured): valid, per-section counts (orgs, contacts, templates, prep, keyDates, orgKinds, ladderRungs), hasVocab, hasPlanner, and issues[] with path (e.g. orgs[3].name), message, severity.
list_workspaces · read-only
Lists the caller's Calaf workspaces — key, display name, position. Use the key as the workspace argument to import_seed.
No parameters. Returns: workspaces[] with key, name, emoji.
import_seed · write (non-destructive, idempotent)
Imports a valid seed into one of the caller's workspaces — the same landing the app's import performs:
- Organizations — deduped account-wide by name, then joined to the workspace
- Templates — deduped within the workspace
- Prep questions — deduped account-wide
- People — never written directly. Contacts in the file are staged as a reviewable batch in Calaf: the user accepts, trims, or discards them together, and can undo an accepted batch in one click. A person must name an org to be stageable.
Re-importing the same file lands nothing twice; nothing is ever deleted or overwritten.
| Parameter | Type | Required | Description |
|---|
workspace | string | yes | Workspace key, from list_workspaces |
seed | string | object | yes | The complete seed JSON document |
parts | string[] | no | Which parts to process: orgs, templates, prep, contacts (= stage people). Defaults to everything the file carries |
Returns (structured): ok, landed (counts per part), failed[], and staged — people, orgs, buildId, skippedWithoutOrg, skippedInvalid.
export_seed · read-only
Serializes a workspace back into a calaf_seed document — the reverse loop: read the book, improve it, re-import (dedupe means a round trip never doubles anything). Privacy defaults: people, the user's personal notes, and content that arrived from other seeds stay out unless explicitly requested (include_people, include_notes, include_seeded); withheld and ceiling-truncated counts are always reported. No plan gate — data portability never sits behind a paywall.
get_profile_schema / get_firm_profile / submit_firm_profile
Firm research at house depth. get_profile_schema returns the module shapes (overview, hq, leadership, comp, process, culture/recruiting notes) and the sourcing epistemics a submission must meet (MBA-grade comp sources, null beats a guess, verify leadership at the source). get_firm_profile reads current state with per-field provenance so chat edits merge against reality. submit_firm_profile writes sourced modules into the user's own overlay — rendered on their firm page badged as assistant research — under the merge contract: empty fields fill, fields the assistant wrote refresh, fields the user typed are never overwritten (reported as conflicts), and clear removes only assistant-owned fields. Sources are required; claims without them drop wholesale.
log_touchpoint · write
Records an interaction (coffee chat, email, event, reply…) with the app's own automations: contact status ladder (never downward), optional follow-up scheduling (which lands on the user's calendar), optional next-day thank-you task, and waking a Researching firm to Networking on real activity.
Completes a person already in the book. Identity fields (email/phone/LinkedIn/school/title/group) fill only where empty — user values win, differences reported. Relationship facets (warmth 1–5, stage, next follow-up, connection type) set explicitly, because "mark Jane dormant" is the ask itself. A firm link fills only when the person has none. append_note adds a dated, attributed note under — never over — the user's own. Creating people stays behind import_seed's staged review.
add_task · write
A due-dated task, optionally tied to a contact or firm. Tasks, follow-ups, and deadlines are what the user's calendar renders — this is the assistant's path onto it.
Paged summaries (contacts include which detail fields are still empty — worth researching) so the assistant navigates the real book, not its memory of it.
log_firm_work / get_firm_work / remove_firm_work · work log
The deals and work items coffee-chat talking points are made of. log_firm_work adds one to a firm's work log with sourcing facets (where it came from, how solid it is: Confirmed / Reported / Rumored); get_firm_work reads the log back, marking which rows are assistant-logged. remove_firm_work deletes only items the assistant itself wrote — entries the user logged by hand are refused, the same class of guarantee as the staged-people rule.
get_applications · read-only
The recruiting pipeline: each application's firm, program, status, deadline, and next interview date.
schedule_interview · write
Puts an interview round on an application's timeline — and so on the user's calendar. Works forward (scheduled) or backward (completed: true with an outcome). Every save also runs the app's own sync: the application's next-interview date recomputes and its status advances along the workspace's real stage ladder (never downgrading, never reopening a closed outcome).
add_application / update_application · pipeline
add_application opens a pipeline row (firm + program; defaults to Watching; duplicates refused). update_application moves it the way the app's board does — status changes append to the stage history and backfill the applied date — and records outcomes (offer_accepted / offer_declined / rejected / withdrawn). Offered comp figures fill only where the user hasn't entered their own; differences are reported, never applied.
get_tasks / complete_task / snooze_task · the to-do loop
List tasks by scope (open / overdue / today / this week / done), mark one done (never delete), or push a due date out. With add_task, the calendar's whole to-do loop runs from chat.
add_firm / update_firm / add_office · the book
add_firm performs the app's two-layer write — global firm + workspace membership — with import's dedupe posture: an existing firm is joined, never doubled, and assistant-added firms are provenance-marked. update_firm sets the facets you name (status, target, priority) and fills prose only where empty (why/intel notes, tier, location — the user's words always win, differences reported). add_office is additive with duplicate-label refusal.
add_note / get_notes / update_note · the notebook
The assistant's scratchpad in the user's notebook: create notes (optionally linked to a firm or person), list with ownership marks, and edit only notes the assistant itself wrote — the user's notes are refused, same guarantee as everywhere else.
create_workspace / rename_workspace / archive_workspace · containers
Create a new workspace through the app's own path (plan slot caps apply), rename it, or archive/restore it. Archive is reversible by design — nothing here deletes anything.
add_template / get_templates / update_template · outreach templates
Reusable email templates per workspace. Assistant templates are marked and editable; the user's own templates and seeded starters refuse edits.
add_prep_question / get_prep / draft_prep_answer · interview prep
Add questions to the prep bank (deduped, optionally firm-tied), list what's unanswered, and draft answers into empty slots only, marked Draft for the user to rework. A question the user already answered is refused outright — their interview voice is theirs.
get_seed_format and validate_seed, exactly as above — no account, no auth. The format is also served as an MCP resource (https://calaf.ai/seed-format.md, text/markdown).
The people posture
An invented firm wastes an afternoon; an invented person is an email to someone who doesn't exist. So people proposed through this API never enter the book directly: they arrive as a named batch in Calaf's review surface, where the user sees the whole list, accepts or discards in one action, and can undo the batch after accepting. Provenance (which assistant, when, from which seed) is stamped on every accepted row.
Trust & data
- OAuth, never passwords. Connection uses OAuth 2.1 with dynamic client registration; the assistant never sees credentials. Disconnect any assistant from Calaf → Settings → Connected assistants, or from your AI's own settings.
- Your book only. Every read and write runs under the connecting user's own account permissions (database row-level security) — a connected assistant can only ever touch that user's data.
- Additive by design. Imports dedupe; nothing deletes or overwrites.
- Health probes: the account server answers anonymous requests with an HTTP 401 + OAuth challenge, as the MCP authorization spec requires — directory health checkers that don't credit auth challenges may show it as "unhealthy." Real MCP clients negotiate it correctly.
Plain HTTP validator (no MCP required)
curl -X POST https://calaf.ai/api/seed-validate \
-H 'content-type: application/json' \
-d @your-seed.json
Same parser, same verdict shape — for CI steps, scripts, and assistants that can only make HTTP requests.
Links