pretix-agent-mcp

Run a self-hosted pretix instance from an AI agent — without the pretix API token ever reaching the agent, and without giving it arbitrary REST access.

Codex / AI agent          untrusted: may be manipulated via prompt injection
        │
        │ MCP over streamable HTTP + bearer token   (or stdio for same-machine dev)
        ▼
pretix-agent-mcp          trusted: holds the credential, enforces capabilities,
        │                 masks personal data, gates irreversible writes
        │ restricted pretix team token
        ▼
pretix REST API

64 purpose-built tools cover routine event administration: create and configure events, clone last year's edition, manage series dates, products, quotas, prices, vouchers, orders and check-in. The web UI is left for one-time organizer setup — notably payment provider onboarding, which involves credentials an agent must never touch.

"Clone this year's conference for 2027, June 12–14, same tickets, early-bird until March." → the agent clones the event, adjusts dates and presale windows, and hands back a preview link. The event exists in draft. Zero UI visits.

"Take the 2027 conference live." → the agent stages it; you run pretix-agent-mcp approve 3f9a1c on the server. That's the whole ceremony.

What makes it safe

Threat	Mitigation
Prompt-injected agent exfiltrating personal data	PII_MODE=redacted by default: names, emails, addresses and phone numbers are masked in every result, and email addresses are scrubbed from pretix error bodies before they reach the agent. A deployment decision, never a tool parameter.
Prompt-injected agent issuing destructive writes	Irreversible operations return a preview and a handle. A human approves them out of band on the server. A confirm: true parameter would just be set by the attacker.
Prompt-injected agent redirecting money or mail	update_event_settings refuses payment routing (provider enablement, bank details, API keys) and mail routing (mail_bcc, smtp_*) outright. Neither is escalated to approval, because neither belongs to an agent at all — and pretix's own permission for payment settings is too coarse to make the distinction.
Agent reaching beyond its remit	Capability allowlist; disabled tools are not advertised at all. The organizer slug is pinned in config. Every path segment is validated. There is no generic HTTP/API tool.
Unauthorized network client	Bearer token required, constant-time compared. Binds 127.0.0.1 by default; refuses a non-localhost bind without a token.
Credential theft	The pretix token lives only in the server's environment — never in results, errors, logs, audit records or tool parameters.

Defaults are closed: localhost, read-only, personal data masked, no writes. Each relaxation is an explicit config change. See SECURITY.md.

Install

git clone https://github.com/bitcoinaustria/pretix-agent-mcp
cd pretix-agent-mcp
uv pip install -e .          # or: pip install -e .
cp .env.example .env         # then edit it

Run it:

set -a && . ./.env && set +a
pretix-agent-mcp serve

Or with Docker:

docker build -t pretix-agent-mcp .
docker run --rm -p 127.0.0.1:8765:8765 --env-file .env \
  -v "$PWD/state:/state" -e STATE_DB=/state/pending-actions.sqlite3 -e AUDIT_LOG=/state/audit.jsonl \
  -e MCP_HOST=0.0.0.0 pretix-agent-mcp

Check what a configuration exposes before pointing an agent at it:

pretix-agent-mcp tools

Connecting a client

Endpoint: https://pretix-mcp.example.org/mcp (streamable HTTP), with your MCP_BEARER_TOKEN. No bridge or proxy is needed — Codex and Claude Code both speak streamable HTTP with a bearer token directly.

Codex CLI (~/.codex/config.toml):

[mcp_servers.pretix]
url = "https://pretix-mcp.example.org/mcp"
bearer_token_env_var = "PRETIX_MCP_TOKEN"   # Codex stores the variable NAME, reads it at connect time
tool_timeout_sec = 60                       # sales_summary can be slow on large events

Then export PRETIX_MCP_TOKEN=... in the shell that launches Codex, or use the CLI:

codex mcp add pretix --url https://pretix-mcp.example.org/mcp --bearer-token-env-var PRETIX_MCP_TOKEN

Gotchas: the key is url (there is no http_url and no type = "http" — transport is inferred); bearer_token is rejected at load in favour of bearer_token_env_var; static headers go in http_headers = { ... } and cannot be set from codex mcp add.

Claude Code:

claude mcp add --transport http pretix https://pretix-mcp.example.org/mcp --header "Authorization: Bearer $PRETIX_MCP_TOKEN"

Same-machine development uses stdio, where the parent process is the peer and no bearer token applies:

pretix-agent-mcp serve --transport stdio

pretix Hosted (pretix.eu) or self-hosted

Both work — it is the same documented REST API. For pretix Hosted set:

PRETIX_BASE_URL=https://pretix.eu

Two hosted-specific things to know:

• Rate limits. pretix Hosted allows 360 requests per minute per organizer for token authentication and answers a 429 with Retry-After; self-hosted instances do not rate-limit by default. The client honours Retry-After and retries up to three times, so a burst degrades into a slower call rather than an error. sales_summary is the tool most likely to hit the limit on a big event (one request per 100 orders scanned) — lower SALES_SCAN_CAP, or ask pretix support for a higher limit, if you run into it.
• Plugins. set_event_plugins can only enable what the instance actually ships; hosted and self-hosted offer different plugin sets. Read the current list with get_event first.

Nothing else differs: team tokens, the event and order APIs, and the settings API behave the same, and the shop URLs the tools hand back are correct for either.

The pretix token

Create a team in pretix (Organizer → Teams) with only the permissions your enabled capabilities need, then add a token to that team. Never use a team that can change teams and permissions.

pretix 2026.x names permissions group:action; older versions show the same things as Can view orders-style checkboxes. Both are listed below — the identifiers are what a 2026.x instance stores, verified against a local 2026.7.

You enable	pretix team permission	older label
read	event.orders:read, event.vouchers:read	Can view orders, Can view vouchers
write (events, settings)	organizer.events:create, event.settings.general:write	Can create events, Can change event settings
write (catalog)	event.items:write, event.settings.tax:write	Can change product settings
write (series dates)	event.subevents:write	Can change event settings
write (vouchers)	event.vouchers:write	Can change vouchers
write (orders: mark paid, comment, resend, edit attendee)	event.orders:write	Can change orders
write (check-in lists)	event.orders:checkin	Can change orders
write:high-risk (cancel, refund, delete)	event.orders:write, event:cancel	Can change orders

Never grant organizer.teams:write (Can change teams and permissions) — that is the one permission that would let a prompt-injected agent widen its own access. Do not grant event.settings.payment:write either unless you accept an agent editing payment configuration; nothing in the tool surface needs it.

Restrict the team to specific events where you can, and set PRETIX_EVENT_ALLOWLIST as a second, server-side limit.

Capabilities

Every tool belongs to exactly one class. MCP_CAPABILITIES decides which classes exist; unlisted tools are never advertised to the client.

Class	Behaviour
read	Executes directly.
write	Executes directly when enabled.
write:high-risk	Records a pending action and returns a preview. Mutates nothing until a human approves.

The live-event guard. A write that targets a live event and can affect price, availability, product structure or existing orders is escalated to write:high-risk at call time. The same operation on a draft or test-mode event executes directly. So agents build and reconfigure drafts without friction, while anything touching a selling event or a customer's money goes through approval.

publish_event is always high-risk — it is the boundary crossing itself. It sets live=true and leaves test mode, because test mode, not the live flag, is what keeps orders from being real. No other tool can leave test mode, so an agent cannot start real sales without an approval.

An operator who accepts the risk can reclassify individual high-risk tools to plain write with MCP_AUTO_APPROVE (e.g. auto-approve publish_event, keep the gate on refunds). That is an explicit, audited, per-tool decision.

The approval ceremony

$ pretix-agent-mcp pending

3f9a1c  publish_event  (expires in 812s)
    Take event 'conf27' (Conf 27) LIVE — the public shop opens.
      starts:  2027-06-12T09:00:00+02:00
      presale: 2027-01-15T00:00:00+01:00 → 2027-06-01T00:00:00+02:00
      test mode: True

approve with: pretix-agent-mcp approve <id>   (1 pending)

$ pretix-agent-mcp approve 3f9a1c
approved 3f9a1c (publish_event)
the agent can now call execute_pending_action with this id

approve --run executes it immediately instead of waiting for the agent. reject <id> throws it away. Nothing here is reachable from the agent: the approval lives on the server on purpose, because a chat-based confirmation is forgeable by a prompt-injected agent and a shell command on the server is not.

Tools

Domain	read	write	write:high-risk
Events	list_events, get_event, get_event_settings, list_tax_rules	create_event, clone_event, update_event¹, update_event_settings¹, set_event_plugins¹, unpublish_event¹, create_tax_rule¹, update_tax_rule¹	publish_event, delete_event, delete_tax_rule
Series dates	list_subevents, get_subevent	create_subevents¹, update_subevent¹	delete_subevent
Catalog	list_products, get_product, list_quotas, get_availability, list_categories, list_questions	create_product¹, update_product¹, create_product_variation¹, update_product_variation¹, create_category¹, update_category¹, create_quota¹, update_quota¹, create_question¹, update_question¹	delete_product, delete_quota, delete_category, delete_question
Orders	search_orders, get_order, search_attendees, sales_summary	mark_order_paid¹, extend_payment_deadline¹, add_order_comment, edit_attendee¹, resend_order_email	cancel_order, refund_order
Vouchers	list_vouchers	create_voucher, create_vouchers_batch¹	delete_voucher
Check-in & waiting list	list_checkin_lists, list_checkins, list_waiting_list	create_checkin_list, update_checkin_list, send_waiting_list_voucher	delete_checkin_list
Approval	get_pending_action	execute_pending_action	—

¹ subject to the live-event guard. A single voucher is an ordinary write; a bulk batch is guarded, because 500 free or quota-blocking vouchers against a selling event is a price and availability change in all but name (a documented tightening of the PRD, which classed all voucher tools as plain write).

Be clear-eyed about what that guard is and is not: it makes the bulk path visible, and an agent that calls create_voucher in a loop reaches the same place without an approval. The guard is not a rate limit, and no per-call gate can be one. Granting write on a live event means trusting the agent with unlimited single vouchers; the audit log is what makes that reviewable afterwards. Deployments that do not want the trade-off keep the voucher tools out of MCP_TOOL_ALLOWLIST, or add create_voucher to a high-risk classification of their own.

Notes:

• sales_summary is computed server-side by paginating orders (pretix has no aggregate endpoint). It returns totals and counts only, in either PII mode, and reports its scan window — large events are slow and the window is capped (SALES_SCAN_CAP).
• All list tools paginate, cap results server-side (≤ 50 per page by default) and mark a truncated answer as such.
• create_subevents is a batch: "every first Thursday until December, 40 seats each" is one call. The agent computes the dates; the server creates them.

Configuration

Environment variables, or a JSON config file passed with --config (env wins). Full list with comments in .env.example.

Variable	Default	Meaning
PRETIX_BASE_URL	—	Base URL of your pretix instance
PRETIX_API_TOKEN	—	Restricted team token. Server-side only
PRETIX_ORGANIZER	—	Organizer slug, pinned; agents cannot change it
PRETIX_EVENT_ALLOWLIST	all events	Comma-separated event slugs
MCP_BEARER_TOKEN	—	Client credential (≥ 24 chars). Required for HTTP
MCP_CAPABILITIES	read	read, write, write:high-risk
MCP_TOOL_ALLOWLIST	—	Expose exactly these tools
MCP_AUTO_APPROVE	—	High-risk tools reclassified to write
PII_MODE	redacted	redacted or full
MCP_HOST / MCP_PORT	127.0.0.1 / 8765	Bind address
AUDIT_LOG	audit.jsonl	Append-only audit log
STATE_DB	pending-actions.sqlite3	Pending approvals
APPROVAL_TTL_SECONDS	900	How long a proposal stays approvable
SALES_SCAN_CAP	5000	sales_summary scan window

Deployment

Put TLS in front of it. Nginx:

location /mcp {
    proxy_pass http://127.0.0.1:8765/mcp;
    proxy_http_version 1.1;
    proxy_set_header Host $host;
    proxy_set_header X-Forwarded-For $remote_addr;
    proxy_buffering off;              # streamable HTTP may hold the response open
    proxy_read_timeout 300s;
}

Keep the bind on 127.0.0.1 and let the proxy be the only route in. If you must bind an external interface, the server refuses to start without MCP_BEARER_TOKEN.

Alongside a self-hosted pretix container

docker-compose.example.yml runs this as its own container next to pretix/standalone, reaching pretix over the internal network (PRETIX_BASE_URL=http://pretix:80).

Do not extend the pretix image to run both in one container, even though pip install pretix-agent-mcp into a derived image would work. The pretix container holds the database credentials, the Django secret key and the ticket data; this one holds an API token limited to a restricted team. Separating them is what keeps a compromise of the agent-facing service — the component deliberately exposed to an untrusted agent — from becoming a compromise of the whole instance, which is the difference between "the attacker can do what that token allows" and "the attacker reads your database". It also keeps docker pull an upgrade instead of an image rebuild, and keeps pretix's own supervisor config untouched.

Two things to keep in mind for that setup:

• Mount a volume for /state. Without it, an approved-but-not-executed action and the whole audit trail vanish on restart.
• docker exec into the MCP container is enough to approve a pending action, so access to the Docker socket is equivalent to approval rights. That is the same trust level as shell access on the host, which is what the out-of-band approval assumes.

Run the server on a different machine from the agent when you can — that is what makes "the credential is not on the agent's machine" a structural property rather than a habit.

Protocol

MCP revision 2026-07-28 via the official Python SDK, which keeps backward compatibility with the initialization-based revisions (2025-06-18 / 2025-11-25). The 2026-07-28 protocol is stateless: no initialize handshake, no session id, protocol version per request. server/discover and the tools/list freshness hints (ttlMs/cacheScope) come from the SDK; cross-call state — pending-action ids — is a server-minted handle passed as an ordinary tool argument.

Authorization deviates from the spec's OAuth 2.1 framework (which is optional): a static bearer token, constant-time compared, appropriate for a single-operator deployment where the operator controls both ends. The spec's token rules still hold — header only, never the query string; missing or invalid token gets 401. Multi-user access would mean implementing RFC 9728 metadata and an external authorization server; that is not built.

Deprecated MCP features are deliberately absent: Roots, Sampling, Logging, HTTP+SSE. Diagnostics go to stderr; writes go to the audit log.

Audit log

One JSONL record per write and per high-risk lifecycle event (proposed / approved / executed / failed), with the tool, PII-redacted arguments, outcome and pending-action id. Reads are not logged. Neither token ever appears.

{"ts":"2027-01-15T10:22:31Z","event":"executed","tool":"publish_event","outcome":"ok","pending_action_id":"3f9a1c","args":{"event":"conf27"}}

Known limits

These are pretix API limits, not workarounds waiting to happen:

• Payment provider onboarding (Stripe/PayPal credentials, OAuth connects) is organizer-level, partly outside the REST API, and involves credentials agents must not handle. One-time UI task.
• Event settings expose the subset the pretix settings API exposes. Where the UI has more, the tools have less.
• Mail routing settings (mail_bcc, mail_from, mail_reply_to, smtp_*) are refused by update_event_settings on purpose: an agent that can BCC every customer mail to an address of its choosing exfiltrates personal data that redaction never sees, because it never passes through the agent's context. Change those in the UI.
• Question options are not editable via PATCH on a question; recreate the question, or edit choices in the UI.
• Payment is configured once, by you, at organizer level. pretix refuses live=true while any product costs money and no payment provider is enabled. Set the provider up on the organizer (its plugin, then its account details) and every event inherits it — including the ones an agent creates later, so paid events publish without anyone opening the web UI again. publish_event's preview lists the products, priced products and quotas it can see, plus a reminder about this precondition; it deliberately does not claim which providers are enabled, because pretix's settings API exposes core settings only and a correctly configured bank transfer or Stripe is invisible there. A free event (a meetup, a stammtisch) needs none of this.
• set_event_plugins can only enable plugins the organizer already allows. Allowing a new plugin is an organizer-level operator task, deliberately outside the tool surface.
• Payment settings are not agent-writable, and that is enforced here rather than by pretix. pretix has one coarse permission (event.settings.payment:write) covering payment deadlines and destination IBANs alike, so an operator who wants an agent to set a deadline would otherwise have to grant it the ability to redirect every payment. update_event_settings refuses the money-routing subset — provider enablement, bank details, anything whose name looks like a credential or an account — while letting payment_term_* through. Third-party providers (BTCPay, Stripe, PayPal) need no special handling: the refusal is keyed on the shape of the setting name, not a list of providers.
• Amounts take at most two decimal places and must be decimal strings — a float is refused rather than rounded, because a rounded price is a rounding bug charged to a customer. Currencies with three decimal places (KWD, BHD, TND) are therefore not supported; widen PRICE_RE in validate.py if you need one.
• Checking someone in is deliberately not a tool: scanning tickets is a physical-presence operation.
• Organizer-level team and token management is out of scope by design — agents must not manage their own permissions.

Development

make setup
make ci      # ruff check + format check + pytest, the same thing CI runs

• AGENTS.md — the working rules (and what Claude, Codex and friends load)
• ARCHITECTURE.md — where the seams are and why
• TESTING.md — the fake-pretix harness and what a change owes in tests
• CONTRIBUTING.md — how to open a PR
• PRD.md — the product spec this was built from

MIT licensed. Contributions welcome — changes to validation, redaction, the approval gate or auth need tests.