MCP

Agisse runs a remote MCP server so an AI client (Claude Desktop, Claude Code, ...) can read and act on one signed-in user's data — the same account as the web app, no separate API key.

Connect

• Endpoint: https://agisse.xyz/mcp
• Auth: OAuth. Your client discovers the flow itself via the standard metadata Better Auth serves at /.well-known/*; when it prompts, sign in with your Agisse account.
• Your account must be approved (or the admin account) — a pending account gets a 403 with This account is not approved for MCP access.

Example client config (Claude Desktop / Claude Code, ~/.claude.json or the equivalent MCP settings):

{
  "mcpServers": {
    "agisse": {
      "url": "https://agisse.xyz/mcp"
    }
  }
}

Connect agent

Don't want to edit config by hand? Paste this into your AI agent (Claude Code, Cursor, Codex, ...) and let it do the setup:

Connect me to the Agisse MCP server.

- Endpoint: https://agisse.xyz/mcp (remote server, streamable HTTP transport)
- Auth: OAuth 2.1 with dynamic client registration. Discovery metadata is at
  https://agisse.xyz/.well-known/oauth-protected-resource and
  https://agisse.xyz/.well-known/oauth-authorization-server. No API key exists;
  do not ask me for one.

Add a server named "agisse" pointing at that URL to your MCP configuration,
using your own CLI or config file (for Claude Code: `claude mcp add --transport
http agisse https://agisse.xyz/mcp`). Then start the OAuth flow and give me the
sign-in link. I sign in with my Agisse account and confirm the consent screen.

When connected, call `list_board` to confirm it works and show me the result.
If you get a 403 "This account is not approved for MCP access", stop and tell
me: my account is still pending approval, there is nothing to fix on your side.

Tools

Tool	What it does
get_profile	Your stable account id, name and email. Read-only.
get_usage	This month's AI budget (EUR): used, limit, remaining, reset date. Read-only.
list_board	List your tracked job applications (Applications), newest first. Paged; filters status, company, since/until (created). Read-only.
get_board_item	Get one job application in full (evaluation, CV, cover letter, notes). Read-only.
list_searches	List your people searches (on Contacts), newest first. Paged; filters company, since/until (created). Read-only.
get_search	Get one search (a company and role) and the ranked people found for it. Read-only.
get_contact	Get one person, including their outreach draft and reply status. Read-only.
list_contacted	List the people you've marked as contacted, newest first. Paged; filters status (reply status), company, since/until (contacted). Read-only.
list_due_followups	List contacted people and applied jobs due a follow-up today (7-day cadence). Read-only.
draft_outreach	Draft the first message for one contact (rewrite: true replaces the stored draft). Sends data to an external AI model and costs from your monthly AI budget; returns the budget before and after, the character limit it enforced (charLimit) and what it drew on (research: no live web research, only facts stored on the contact). Only drafts — never sends anything. An existing draft without rewrite is returned free.
update_contact_status	Update a contact's reply status or mark a follow-up sent. Optional expectedStatus makes it a no-op with an error if the status has changed. Returns { before, after }.
get_seeker_brief	Your CV and target industries, for research and company ideation done by your own AI. With jdText, also the evaluation rubric and the exact output schema add_job expects. Read-only, free.
add_job	Save a posting to Jobs with an evaluation your own AI produced (jdText, optional url, evaluation). No model call, nothing charged. A posting already in Jobs comes back with existing: true and nothing is written.
add_to_shortlist	Save a company to target (name, optional domain, why). Saving the same name again updates it. No model call.
list_shortlist	List the companies on your shortlist, newest first, with why. Read-only.
list_vacancies	Your Jobs for you feed: open early-career postings in Europe, ranked for you, 30 per page (cursor / nextCursor). Each has title, company, url, location, fit (0–100), reasons (Jev's verdict first when there is one) and deadline. Filters (q, countries, level, functions, industries, languages, remoteOnly, deadlineWithin, sort) default to your saved feed filters. Read-only, no model call.
update_job_status	Move a job to another status (applied, interview, offer, …), as on the web. Optional expectedStatus makes it a no-op with an error if the status has changed. Returns { before, after }.

Every tool declares an outputSchema and returns its result both as JSON text and as structuredContent; errors come back as isError text.

Paging

The three list_* tools take limit (1–100, default 50), cursor, company (case-insensitive substring), since (inclusive) and until (exclusive) as ISO dates or datetimes in UTC, and return { items, nextCursor, total }. total counts everything matching the filters; pass nextCursor back unchanged for the next page, until it is null.

Use your own AI

The web app evaluates postings and scouts companies with Agisse's own models, paid from your monthly AI budget. Over MCP you can do that work in the client instead, on whatever subscription it runs on, and only hand Agisse the result:

1. get_seeker_brief gives your CV and targets. Ask your AI to research companies or roles with its own web search and to shortlist the good ones with add_to_shortlist.
2. For a posting, pass its text to get_seeker_brief: you get the same rubric and output schema the web app uses. Your AI evaluates the posting by that rubric.
3. add_job stores the evaluation. The job shows up in Jobs and Applications like any other, with score and recommendation, and costs nothing from the budget.
4. Discover → track: list_vacancies shows your feed, your AI picks the roles worth it, evaluates them and saves them with add_job, then update_job_status moves each one along as you apply, hear back and interview.

The server makes no model call for any of these tools. draft_outreach is the only MCP tool that spends budget.