Skip to main content
Glama

standup-mr

standup-mr MCP server Listed on mcpservers.org mcprush grade

Standup notes from merge request state, not commit logs.

Most standup tools read your local git log. That answers "what did I type", which is not what anyone asks in a standup. This one reads GitLab or GitHub: what is ready to merge, what is blocked, what is waiting on you — and when a pipeline is red, it opens the job log and tells you why.

standup-mr reading a GitLab project: ready, blocked, draft, and the error line from a failed job

What makes it different

commit-log tools

standup-mr

Source

local git log

GitLab or GitHub API

Merge request state

✗

ready / blocked / draft / stale

Review queue

✗

pending only, approvals filtered out

Failed pipeline

✗

error lines from the job trace or the Actions job log

Self-hosted (GitLab CE/EE, GitHub Enterprise)

varies

first class

Related MCP server: Slack Gitlab Digest Model Context Protocol MCP Server

Use

npx standup-mr fetch                            # JSON, provider auto-detected
npx standup-mr fetch --provider github          # GitHub
npx standup-mr fetch --provider gitlab          # GitLab
npx standup-mr fetch --markdown                 # structured digest
npx standup-mr fetch --lang tr                  # Turkish date labels
npx standup-mr fetch --markdown | npx standup-mr post --google-chat "$URL"

Identity

GitHub

GitLab

Flags

--host / --token

--host / --token

Env

GITHUB_HOST / GITHUB_TOKEN

GITLAB_HOST / GITLAB_TOKEN

CLI session

gh auth config

glab auth config

GitHub defaults to github.com when no host is given. GitLab has no default — self-hosted is the norm there, so a host must come from a flag, env var, or glab's own config.

The provider itself is picked in this order:

  1. --provider github / --provider gitlab, if passed

  2. a recognizable --host (github.com, gitlab.com, or a hostname containing github/gitlab)

  3. STANDUP_PROVIDER, or whichever of the GITHUB_* / GITLAB_* env pairs is set

  4. whichever of gh / glab is logged in

If none of these resolve — or both do, ambiguously — the command fails with a clear error instead of guessing.

So if you already use gh or glab, there is nothing to configure.

Post it to chat

npx standup-mr fetch --markdown | npx standup-mr post --slack "$SLACK_WEBHOOK_URL"

Hosted connector: any AI assistant

No token, no config file — sign in instead. The hosted server speaks the standard MCP OAuth flow, so any assistant that supports remote MCP servers can use it: add the URL, and a browser page opens where you sign in.

https://standup.ilkerbalcilar.com/mcp

Add to Cursor Install in VS Code

Assistant

How

claude.ai (web, mobile, Desktop)

Settings → Connectors → Add custom connector → paste the URL → Connect

Claude Code

claude mcp add --scope user --transport http standup https://standup.ilkerbalcilar.com/mcp, then /mcp to sign in

Cursor

The button above, or "standup": { "url": "https://standup.ilkerbalcilar.com/mcp" } in ~/.cursor/mcp.json

VS Code (Copilot)

The button above, or "standup": { "type": "http", "url": "https://standup.ilkerbalcilar.com/mcp" } in mcp.json

Gemini CLI

gemini extensions install https://github.com/Jubstaaa/standup-mr, then /mcp auth standup

Codex CLI

codex mcp add standup --url https://standup.ilkerbalcilar.com/mcp, then codex mcp login standup

ChatGPT

With developer mode on: Settings → Apps → Create → paste the URL, OAuth

claude.ai and Claude Code are tested end to end; the others use the same standard flow. If one of them doesn't sign in, please open an issue.

On the sign-in page you link the accounts the note should read:

  • GitHub — sign in through the standup-mr GitHub App. It only sees the repositories it is installed on, so install it on your org too; if the org won't approve it, add github.com with a fine-grained token instead.

  • gitlab.com — sign in with GitLab.

  • Your company's GitLab or GitHub Enterprise — type the host. The page shows the exact redirect URI and scopes for an OAuth application on your host; you paste its ID and secret and sign in with it. The application is kept for your connection only, so register your own rather than sharing a colleague's. A read-only token works too.

  • Chat webhook (optional) — Slack, Discord or Google Chat, so your assistant can post the finished note when you ask it to.

Link as many as you like — say, GitHub for side projects and your company's GitLab for work — then ask for "my standup" or "my work standup". With several accounts linked, the assistant asks which one you mean.

Added in an assistant with a mobile app — Claude or ChatGPT — it works from your phone too: ask for your standup on the way to the meeting.

What is stored. Your provider tokens, webhook URL and, for a self-hosted host, your OAuth application's ID and secret, encrypted with AES-256-GCM, and only to read your merge requests and post your note. The access tokens your assistant holds are stored as hashes. Disconnecting deletes all of it; connections unused for 30 days are deleted. One exception for now: self-hosted OAuth applications registered before per-connection apps were introduced are also kept in the old shared table, so a rollback still works, until the next release removes it. The server is the code in remote/.

When to use the local server instead. A GitLab behind a VPN or on a private network is out of reach of a hosted server. Run the stdio server below on your own machine — the token never leaves it.

The four surfaces

CLI — the core. Emits JSON; one runtime dependency (p-map).

MCP server (mcp/) — three tools for Claude Desktop, Claude Code, Cursor, or any MCP client that runs a local process. See mcp/README.md.

Remote MCP server (remote/) — the same tools over HTTPS with sign-in, for any assistant that supports remote MCP servers, including the web and mobile apps. See above.

Claude Code plugin — the note-writing playbook, shipped as the standup skill. From inside Claude Code:

/plugin marketplace add Jubstaaa/standup-mr
/plugin install standup@standup-mr

Then type /standup. Updates come with /plugin marketplace update standup-mr.

Using it from Cursor, Codex, or another assistant

If you're not using Claude Code, wire up the MCP server for live data and paste in the note-writing rules separately.

Cursor — add to ~/.cursor/mcp.json (global) or .cursor/mcp.json (project-local):

{
    "mcpServers": {
        "standup": {
            "command": "npx",
            "args": ["-y", "standup-mr", "mcp"],
            "env": {
                "GITHUB_TOKEN": "ghp_..."
            }
        }
    }
}

Codex — add to ~/.codex/config.toml:

[mcp_servers.standup]
command = "npx"
args = ["-y", "standup-mr", "mcp"]

[mcp_servers.standup.env]
GITHUB_TOKEN = "ghp_..."

The exact config key and file path are version-dependent for both clients — if a snippet above doesn't work, check Cursor's MCP docs or Codex's own config documentation for the current format rather than trusting this file blindly.

The server exposes three tools:

Tool

What it does

get_standup_data

Reads the provider and returns the report as JSON. Optional provider, host, lang.

get_note_instructions

Returns the note-writing rules, so the assistant can write the note the way the skill would.

post_standup_note

Posts a finished note to a Slack, Discord or Google Chat webhook.

post_standup_note reads the webhook URL from STANDUP_WEBHOOK_URL and never takes it as an argument — anyone holding that URL can post to the channel, so it belongs with the tokens, not in a transcript. The payload shape is inferred from the URL: an https webhook path on hooks.slack.com, chat.googleapis.com or discord.com, with no credentials or custom port. Any other URL needs kind.

Three shapes are implemented: slack and google-chat both post {"text"}, discord posts {"content"}.

Slack and Google Chat do not render standard Markdown, so the note is rewritten on the way out: **bold** becomes *bold* and a ## heading becomes a bold line. Inline code, fenced blocks and their contents are left alone. Discord speaks Markdown natively and is sent untouched. Mentions are neutralised on every kind, so a merge request title holding <!channel>, <@U1> or <users/all> cannot ping anyone. Anything else that accepts a Slack-shaped body — Mattermost, Rocket.Chat, an n8n or Zapier endpoint — works today by passing kind: "slack", or --slack URL on the CLI.

Outside MCP, the same rules are available on stdout:

npx standup-mr instructions >> AGENTS.md

--markdown is a digest, not a written note

--markdown organizes the raw material into readable sections. It does not group events into themes or diagnose blockers — that is the model's work, and it lives in the MCP client's prompt or in the Claude Code skill.

  • Without AI: a structured digest.

  • With AI: a note you can read out.

What the digest looks like

Anonymised output from a real Monday run — note that Friday and Saturday each get their own section, and that a merge request GitLab has not evaluated is not called ready:

# Monday, 31 August — dev

_Structured digest — not a written note._

## Previous working day: Friday, 28 August

- `acme/ui` pushed to — fix(keyboard): scale keys to viewport (4 commits)
- `acme/ui` accepted — chore(deps): bump @acme/ui to 0.5.18
- `acme/api` opened — feat: package subscription sales

## Previous working day: Saturday, 29 August

- `acme/ui` pushed to — fix(keyboard): close the autofill bar (1 commit)

## Ready to merge (2)

- `acme/api` !196 fix: normalise the +90 trunk prefix
- `acme/web` !194 refactor: loading state — **no pipeline ran**

## Blocked (1)

- `acme/terminal` !49 fix: relative date chips — **1 unresolved comment(s)**

## Reviews (2 pending)

- `acme/mobile` !501 chore: upgrade to RN 0.87 — Teammate

## Blockers

- `acme/mobile` !6 — job `quality`
    - `npm ERR! code E404`
    - `npm ERR! 404 Not Found - GET https://registry.example.com/@acme%2fui`

The last section is the point of the tool. Every other standup tool can tell you that pipeline is red; this one opens the failed job's log and shows you the 404 — and a 404 rather than a 403 usually means the token's scope is wrong, not that the package is missing.

Stability

Since 1.0.0 the CLI commands and flags, the JSON that fetch and get_standup_data return, the MCP tool names and arguments, and the environment variables follow semantic versioning: a change that would break an existing setup waits for a major release. New fields in the JSON may appear in minor releases, so read it by name rather than by position. The hosted server at standup.ilkerbalcilar.com is updated continuously and is not tied to the npm version.

Known limits

  • GitHub's events feed is shallow. It is capped at roughly 300 events over the last 90 days, so a very active account or an old gap can silently lose the earliest events. GitLab has no comparable documented cap.

  • GitHub activity is only visible to a token belonging to that same account, and private-repository events do not show up for anyone else's token, even with otherwise sufficient scopes.

  • On GitHub, CI that reports only through the legacy commit-statuses API — still how some vendors integrate — shows up as pipelineMissing. Check state is read from check-runs only.

  • A blocker whose diagnosis could not be fetched is still reported, as job: "unknown" with a diagnosis unavailable: …, job log unavailable: HTTP N or no failed job found … error line. The merge request is blocked either way; only the explanation is missing. Server errors are retried twice first, and a rejected token still fails the run.

Upgrading from 0.1.x

0.2.0 changes the JSON output, the library API, and the MCP options. If you pipe standup fetch into anything, or import the package, read CHANGELOG.md before upgrading. The short version:

  • previous and previousEvents are replaced by previousDays[], one entry per active day, so a weekend no longer swallows Friday. jq .previous now returns null with no error.

  • MergeRequest, Review and Blocker carry a required provider field, and Provider.getReviews takes an Identity rather than a numeric id.

  • The MCP CollectOptions.provider is now a provider name; inject a Provider instance through providerImpl.

Claude Code plugin users should run /plugin marketplace update standup-mr — the standup skill changed along with the report shape.

Requirements

Node 22 or newer. The CLI core (fetch, post, instructions) loads one runtime dependency, p-map. The MCP server (mcp command) also loads @modelcontextprotocol/sdk and zod. All three are installed with the package.

License

MIT

Available Tools

3 tools
get_note_instructionsGet note instructionsA
Read-onlyIdempotent
Inspect

Returns the note-writing rules — the playbook for turning get_standup_data output into a standup note a person would actually say out loud: how to group the previous day by theme, how to derive today from open merge requests, and how to report a blocker from its job log. Read-only, no arguments, no network. Call it once before writing the first note; the rules do not change between calls.

Inputs: no named parameters.

get_standup_dataGet standup dataA
Read-onlyIdempotent
Inspect

Collect merge-request-based standup data from GitLab or GitHub. Returns one JSON object: today, previousDays[] with the events of each, todayEvents[], myMrs[] bucketed ready / blocked / draft / stale, reviews[] waiting on you, and blockers[] carrying the error lines read out of each failed pipeline job log.

Call it once at the start of a working day, to write a standup note. It is a snapshot, not a search API: it cannot fetch one named merge request, reach further back than the previous working day, or filter by project.

Read-only, and credentials never come from an argument — they come from the environment or a logged-in gh / glab session. A rejected token, a refused resource or a rate limit fails the call with the host's own message, after two retries on transient server errors. A blocker whose diagnosis could not be fetched is still returned, with job: "unknown", so a red pipeline is never silently dropped.

Inputs: host, lang, provider.

post_standup_notePost standup noteAInspect

Posts a finished standup note to a chat webhook. This one has a side effect: it sends a message other people will see, so only call it on a note the user has agreed to send.

The webhook URL is read from STANDUP_WEBHOOK_URL, never from an argument — a webhook URL is a credential, since anyone holding it can post to the channel. Slack and Discord payload shapes are supported; the shape is inferred from the URL host, and kind is only needed for a proxied or self-hosted endpoint whose host gives nothing away — Mattermost and Rocket.Chat take the slack shape. A missing URL, an unrecognised host, or a webhook that rejects the message comes back as an error rather than a silent success.

Inputs: text (required), kind.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.4/5.0

    Scored across 3 tools

    Disambiguation5/5

    Each tool maps to a distinct step: fetching standup data, fetching writing rules, and posting a note. The descriptions explicitly bound each tool's scope and note what it cannot do, so an agent cannot easily confuse them.

    Naming Consistency5/5

    All three follow the same verb_noun snake_case pattern (get_standup_data, get_note_instructions, post_standup_note). The get_/post_ split also correctly signals read-only versus side-effecting tools.

    Tool Count4/5

    Three tools is lean but every one earns its place in the fetch-rules-post workflow, with no redundant operations. The surface is deliberately narrow, which suits a single-purpose standup server but leaves little room for variation.

    Completeness4/5

    The core lifecycle (gather data, apply rules, publish note) is fully covered, including error/blocker reporting. Gaps are documented by design — no single-MR lookup, no date-range or project filtering, and no way to preview or amend a posted note — all of which an agent can work around.

    Maintenance

    ActivityActive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers