Instagram Analytics MCP

An MCP server that answers questions about Instagram reel performance in plain language, across multiple accounts.

The point is not that it wraps an API. It is that the number that actually predicts reach does not exist in the Instagram API, so the server computes it.

"How did my last 10 reels do?"
"What worked best this month?"
"Which of my accounts is working?"

The problem this solves

Instagram's Graph API returns views, reach, saves, shares and average watch time.

It does not return completion rate -- the share of the video people actually watch. On a real account measured over ~700 reels, completion is what separates a reel that dies from one that travels:

Views are the outcome. Completion is the cause, and it is readable within hours of posting rather than days.

Computing it needs avg_watch_time / duration. Duration is not in the API either. So the server probes each video's media_url with ffprobe to measure it.

That is the whole reason this exists. Two hops the API will not do for you, plus a threshold judgement the API has no opinion about.

Tools

Every reel comes back with date, completion, a verdict label, duration, views, reach, saves, shares, the caption's first line as the hook, and a permalink.

Works with one account or several. account is optional and defaults to the first one you configured.

Requirements

• An Instagram Professional account (Business or Creator). Personal accounts cannot use the Instagram API at all

• Python 3.10+

• ffprobe (brew install ffmpeg) -- without it there is no duration, so no completion rate

Setup

git clone https://github.com/sskghub/instagram-analytics-mcp
cd instagram-analytics-mcp

python3 -m venv .venv
.venv/bin/pip install -r requirements.txt

cp .env.example .env

Then get a token. SETUP.md is the full walkthrough, about 15 minutes for the first account: create a Meta app, add Instagram, generate a token.

Once a token is in .env, this checks everything and tells you the account id to paste back in, so you never have to hunt for it:

.venv/bin/python check_setup.py

[  OK  ] mcp package installed
[  OK  ] ffprobe found
[  OK  ] main: token works, account @yourhandle

Then confirm it pulls real data, and that the MCP layer works end to end:

.venv/bin/python server.py --selftest
.venv/bin/python test_server.py

Register with Claude Code:

claude mcp add ig-analytics -- /absolute/path/.venv/bin/python /absolute/path/server.py

The server reads its own .env, so no credentials go in the MCP config file. That config is committed; tokens are not.

Adding another account means adding two lines to .env. There is no code to edit -- accounts are discovered from the variable names.

Token expiry

Instagram tokens last ~60 days. When one dies, everything downstream silently returns nothing.

refresh_tokens.py exchanges a still-valid token for a fresh 60-day one:

python refresh_tokens.py --if-older-than 7

Run it weekly. The constraint that shapes the design: an expired token cannot be refreshed. Meta will not renew a dead token, so refreshing early is the only strategy that works. Each refresh resets the full 60 days, so early refreshes cost nothing.

Scheduling notes, including the macOS trap where a launchd job silently cannot read your files, are in SETUP.md.

It backs up .env before writing, rewrites duplicate keys, and alerts on failure.

Refreshing does not invalidate the old token, so several machines can each refresh their own .env independently. Token values never need syncing between hosts.

Notes from building it

Things that cost real time, kept here because they are the parts that generalize.

sys.exit() is fine in a CLI and fatal in a server. The first version reused a function from an existing command-line script. That function called sys.exit() when a token was rejected, which would have killed the whole server process on the day a token expired. Tools now raise ValueError; the SDK turns standard exceptions into readable results the model can act on, and the server survives.

Errors should say what to do. A dead token returns the steps to regenerate it, not a stack trace. The model can relay that to a human who can actually fix it.

The docstring is the interface. It is how the model decides whether to call a tool at all, so each one says when to reach for it, not just what it returns.

Scheduled jobs can fail silently. On macOS a launchd timer for the refresh script failed with Operation not permitted, because TCC blocks background agents from reading protected directories. It reported as loaded and would have quietly never run. Forcing a run and reading the log is the only way that surfaces.

Duplicate keys in .env are a real trap. A stale duplicate can shadow a freshly written token depending on how the loader resolves them, so the writer rewrites every occurrence rather than the first.

The API changed names. It is mcp.server.mcpserver.MCPServer; the older mcp.server.fastmcp.FastMCP path was removed in mcp 2.x, along with other legacy modules. Most examples online still show the old import and will not run.

License

MIT