Introducing CSV Output Format in Insiders Mode

[RossTarrant]

Hello from the GitHub MCP Server maintainers 👋

We’ve recently added an experimental CSV output format for supported list-based tools in the default toolset, available now through Insiders Mode.

This is now available on the remote GitHub MCP Server and in the latest local GitHub MCP Server release, v1.1.0.

The goal is to reduce token usage for list-heavy workflows by returning compact CSV instead of JSON, avoiding some of the repeated object and array structure that appears in JSON responses.

Important

This currently applies only to supported list-based tools in the default toolset, such as tools whose names start with list_. It does not apply broadly to every tool or every toolset.

Example token savings

The examples below were captured against the github/github-mcp-server repository and compare the same tool responses in JSON vs CSV.

These figures are approximate. Token counts were calculated from the captured payloads using the cl100k_base tokenizer to provide a consistent comparison point, but actual savings may vary depending on various factors such as the model, MCP host, tool, arguments, result count, and shape of the returned data.

Tool	Results	Approx. JSON tokens	Approx. CSV tokens	Approx. tokens saved	Approx. reduction
list_pull_requests	30	22,561	16,902	5,659	25.1%
list_commits	30	17,805	13,396	4,409	24.8%
list_issues	30	16,120	13,429	2,691	16.7%

These are representative examples only. Your results may be higher or lower depending on the context of the tool call.

How to try it

CSV output is available to MCP Insiders. Insiders Mode lets users opt into experimental GitHub MCP Server features before they reach general availability.

Deployment	How to enable Insiders Mode	Example
Remote server	Add the X-MCP-Insiders: true header	Configure your MCP client to send the header
Remote server quick access	Append /insiders to the remote MCP URL	https://api.githubcopilot.com/mcp/insiders
Local / OSS server	Use the --insiders CLI flag	github-mcp-server stdio --insiders
Local / OSS server	Set the environment variable	GITHUB_INSIDERS=true

For more context, see the Insiders Mode announcement.

We want your feedback

If you’re using Insiders Mode, please try the CSV output with your list-based workflows and let us know:

• Are you seeing meaningful token savings?
• Which list tools benefit the most?
• Does CSV make responses easier or harder for your agent to work with?
• Are there formatting changes or improvements that would make this more useful?

Please leave a comment below with your feedback, examples, or suggestions.

[mnifzied-create]

Nice — output compression is a real lever; the JSON structure tax on list-heavy results adds up fast. For anyone budgeting GitHub MCP token usage, it's worth pairing this with the other half of the ledger: the tool-schema input cost, which is fixed and re-sent every turn regardless of output format.

I measured that across 13 MCP servers recently — the GitHub MCP server's tool definitions total ~3,546 tokens/turn (26 tools), the highest in the set, paid on every request before any result comes back. Fattest single schema was create_pull_request_review (~360 tok). Method + data: https://mnifzied-create.github.io/agentloop/token-tax/ (I used o200k_base, so a few % off your cl100k figures, but the structure holds).

Two things that compound nicely with the CSV work:

• Serialization cruft on the schema side — ~20% of typical schema bytes is non-semantic ($schema, additionalProperties, auto-added title), strippable at zero tool-selection cost. Same spirit as JSON→CSV, but on the input.
• Toolset scoping — since 3,546 is the full default toolset, the per-turn schema cost scales with how many tools are loaded, so list_-tool CSV + a lean default toolset hit both sides at once.

Output size is the half people notice; the per-turn schema overhead is the half they usually don't — great to see the team chipping at the cost either way.