SpritesAdd Sprites to Codex
Available on Anywhere Codex CLI runs — a ChatGPT plan that includes Codex, or an API key. You need a Sprites account with PostHog already connected — the connector acts on that connection, so Codex never handles a PostHog credential.
Add the server
One command writes the entry into ~/.codex/config.toml for you.
Terminalcodex mcp add sprites --url https://mcp.sprites.ai/mcpOr write the config by hand
The same entry in ~/.codex/config.toml, or in .codex/config.toml inside a trusted project when you want the connector scoped to one repo.
~/.codex/config.toml[mcp_servers.sprites] url = "https://mcp.sprites.ai/mcp" auth = "oauth"Sign in
Codex opens a browser for the Sprites consent screen and stores the token locally. Sprites supports dynamic client registration and client ID metadata documents, so there is nothing to pre-register.
Terminalcodex mcp login spritesApprove changes by name
Codex renders results as text, so a pending change comes back written out rather than as a card you can click. Nothing is applied until you approve it, and the approval waits for you in Sprites — open it there and the same run continues.
Pending changeChangeDaily budget $40 → $65Gategate_7f2c…
Official PostHog MCP vs Sprites MCP
PostHog runs its own MCP server at https://mcp.posthog.com/mcp. Here is what each one covers, so you can pick the one that fits in Codex — or connect both.
PostHog hosts its own MCP server, free to call, with a one-line installer (npx @posthog/wizard mcp add) for Claude, Cursor, VS Code, Windsurf, Zed and others. It reaches most of PostHog — insights and SQL, dashboards, experiments, feature flags, surveys and error tracking — and it reads and writes. Query parameters on the URL can narrow it to chosen features or named tools. Tools that use an LLM inside PostHog bill as PostHog AI. See PostHog’s own documentation.
| PostHog officialsince 2025 | Sprites MCPapproval-gated | |
|---|---|---|
| Endpoint | mcp.posthog.com/mcp | mcp.sprites.ai/mcp |
| Scope | Most of PostHog: insights, SQL, flags, experiments, surveys, errors | Product analytics: config, trends, breakdowns, funnels, retention |
| Writes | Yes — flags, insights, dashboards, experiments, issues | None in PostHog; ad changes held as approval cards |
| Ad spend | Not in view | Google Ads, Meta, LinkedIn, TikTok and Reddit in the same thread |
| Figures | Returned for the model to read | Computed in code; totals reconciled against rows |
| Cost | Free; LLM-backed tools bill as PostHog AI | Runs on your Sprites plan credits |
If you work inside PostHog — shipping flags, running experiments, triaging errors — PostHog’s own server is the right one, and its URL filters keep it from flooding your client with tools. Use Sprites when the question starts with spend: which campaign, which ad set, which channel brings the users who stay — and when the answer should turn into a change in the ad account rather than in PostHog.
Why run PostHog ads through Sprites
Codex is the interface; the connector decides what it can see and what it is allowed to do. Six things change when that connector is Sprites.
The click and its cost in one answer
PostHog sees what a visitor did; the ad account sees what the visit cost. Sprites reads both, so "cost per activated user by campaign" is one question.
Every channel on one connectorRead-only on PostHog
Nothing Sprites does can flip a flag, end an experiment or edit a dashboard. The only changes it proposes are in the ad account, one approval card at a time.
Numbers that are checked before they are said
Every figure is computed in code from a query you can see, and totals are reconciled against their rows. A figure that does not reconcile is withheld, not rounded into the answer.
Run the first prompt
Ask Codex this. It routes to a Sprites door, which reads your PostHog account and answers — no account to name, no confirmation to give first.
- Ask Sprites which utm_campaign brought the users who finished onboarding last month.
- With Sprites, show the signup-to-activation funnel for Google Ads visitors versus Meta visitors.
- Use Sprites to check what events our PostHog project actually tracks before we build the report.
- Have Sprites compare 4-week retention for users from paid search against organic.
- With Sprites, find the Meta ad sets whose visitors never reach the pricing page.
Run `codex mcp list` and confirm sprites shows as connected.
What Codex can do with PostHog
The same doors, whichever client you are in. Marketing work that belongs in a commit: Codex can read the landing page it is about to ship and brief Sprites on the campaign pointing at it, in one session.
Post-click, by campaign
Break a funnel or a trend down by the session’s entry UTM source, medium, campaign, term or content, its gclid or fbclid, or its channel type. That is the join an ad platform cannot make on its own.
It reads the project before it guesses
Every event name with its 30-day volume and reach, the property keys, saved insights, flags and cohorts. A misspelled event does not come back as a confident zero.
Figures that add up
Totals are queried separately from the rows and reconciled. When unique users cannot be summed across rows, the answer says so instead of reporting a gap that is not there.
Your internal users stay out
The project’s own internal and test-user filter is applied by default, and the answer always says whether one exists.
Read-only, by design
Sprites queries PostHog and changes nothing in it: no flags flipped, no dashboards rewritten, no experiments started.
The ad account is in the same thread
The Google Ads, Meta, LinkedIn, TikTok and Reddit specialists all read PostHog, so "which campaign brings users who activate" is one question, not two tools and a spreadsheet.
How approvals work here
Codex renders tool results as text, so a PostHog change comes back written out rather than as a card. Nothing is applied until you approve it, and the approval waits for you in Sprites — open it there and the same run continues.
What it reads
- Events, their 30-day volume and reach, and property keys
- Saved insights, feature flags, cohorts and actions — listed, not changed
- Trends, breakdowns, funnels and retention computed from your events
- Session attribution: entry UTM parameters, gclid, fbclid, referring domain and channel type
- The project timezone and its internal and test-user filter
What it can change
- Nothing in PostHog. Every PostHog tool is read-only.
- Changes the analysis leads to — a budget, a paused ad set — happen in the ad account, each as an approval card
When it does not work in Codex
Three things go wrong in this client specifically, and each has a one-line fix.
Codex says the tool is not found.
The config was edited while a session was open. Restart Codex — it reads config.toml at startup.
Login loops without ever completing.
The OAuth callback defaults to http://127.0.0.1/callback. If something else on the machine holds that port, set mcp_oauth_callback_url in config.toml and log in again.
Codex says a change was applied and nothing happened in the account.
It was proposed, not applied — that is the design. The approval is waiting in Sprites; open it there and the run continues from where it stopped.
And on the PostHog side
Sprites says PostHog is not connected.
Connect PostHog in Sprites with a personal API key, then select the project. A key with no selected project cannot be queried.
An event comes back as zero.
Event names are matched exactly. Ask Sprites what the project tracks first; it lists every event with its 30-day volume, so a misspelling is obvious.
Sessions do not match any campaign.
The breakdowns read the UTM parameters and click IDs a session arrived with. Ads without UTM tags, or landing pages that strip them, arrive as direct traffic; tag the ads and the join works from the next click.
The numbers include our own team.
The project has no internal and test-user filter. Set one in PostHog’s project settings; Sprites applies it automatically from then on.
PostHog and Codex: questions
Which URL do I paste into Codex?
https://mcp.sprites.ai/mcp. There is one endpoint and every client uses it — every OAuth grant is bound to exactly that address, so a variant of it will not be honoured.
Can Codex change my PostHog account without asking?
Codex renders tool results as text, so a PostHog change comes back written out rather than as a card. Nothing is applied until you approve it, and the approval waits for you in Sprites — open it there and the same run continues.
Do I need a PostHog API key or developer token?
No. PostHog is connected once inside Sprites through PostHog (personal API key and a selected project), and the connector acts on that connection. Codex never sees a credential for it.
Why Codex for PostHog specifically?
Marketing work that belongs in a commit: Codex can read the landing page it is about to ship and brief Sprites on the campaign pointing at it, in one session.
Is there an official PostHog MCP server?
Yes. PostHog runs its own hosted server at https://mcp.posthog.com/mcp, signed in with PostHog OAuth or a personal API key, and installable with npx @posthog/wizard mcp add. It reads and writes across insights, dashboards, SQL, feature flags, experiments, surveys and error tracking. Calling it is free; tools that use an LLM inside PostHog bill as PostHog AI.
What is the difference between that and Sprites?
PostHog’s server is PostHog, all of it, including the ability to change flags and dashboards. Sprites reads PostHog only, and puts it beside your ad accounts: the same specialist that reads Google Ads or Meta reads the funnel, so it can say which campaign brings users who activate — and propose the budget change.
Keep going
The same Codex connector reaches every other channel — nothing else to add.
PostHog, from Codex
Paste one URL, sign in once, and the next PostHog question you ask Codex gets answered from the live account.