Scribe MCP Server
Scribe is a Model Context Protocol (MCP) server that keeps project documentation and progress logs consistent across agents. This repo ships a lightweight stdio server, filesystem/DB storage backends, and tooling to keep docs in sync with day-to-day development.
Prerequisites
Python 3.11+
pip(or your preferred package manager)(Optional) PostgreSQL if you want shared storage instead of the default SQLite backend
Install Python dependencies:
Quick Start
Install dependencies (see above).
Choose a storage backend via environment variables:
SQLite (default): no action required.
PostgreSQL: set
SCRIBE_STORAGE_BACKEND=postgresandSCRIBE_DB_URL=postgresql://....
Launch the server:
python server.py
Claude / MCP Configuration
Copy config/mcp_config.json into your Claude Desktop/Code MCP configuration and adjust the SCRIBE_ROOT path to point to this repository:
Smoke Test
Verify the MCP server responds before wiring it into an IDE:
This script runs a short-lived stdio session and checks that the server advertises tools.
Reminder Metadata
Every MCP tool responds with a reminders array containing context like stale docs, overdue logs, and the active project. Use these cues to keep architecture, phase plan, and checklist in lockstep with the work you’re doing.
Customising reminders
Set
defaults.reminderinside a project config (orSCRIBE_REMINDER_DEFAULTSenv JSON) to tweak behaviour. Example:
Environment variables:
SCRIBE_REMINDER_IDLE_MINUTES— gap (minutes) before a new work session resets and warm-up kicks in (default: 45).SCRIBE_REMINDER_WARMUP_MINUTES— grace period after resuming before warnings escalate (default: 5).SCRIBE_REMINDER_DEFAULTS— JSON blob applied to every project when specific overrides aren’t set.
Project Structure
MCP_SPINE/– container package for MCP servers.scribe_mcp/– Scribe server entrypoint, tools, storage backends, utilities.config/projects/– per-project Scribe configs.config/mcp_config.json– sample MCP configuration for Claude.docs/dev_plans/– architecture, phase plan, checklist, and progress log.scripts/test_mcp_server.py– smoke test for the MCP server.tests/– unit tests (run withpytest MCP_SPINE/tests).
Development Workflow
Select a project with
set_project(either via MCP tool orscripts/scribe.py).Use
append_entryto log every meaningful change—Scribe tracks both the markdown log and (optionally) a database mirror.Keep docs aligned with implementation; reminder tooling surfaces when architecture/phase plan/checklist drift.
Troubleshooting
Missing MCP SDK: install
pip install mcpif the server exits with “MCP Python SDK not installed.”No tools returned: ensure
scribe_mcp/tools/__init__.pyimports all modules and that your virtual environment is active.SQLite permissions: confirm
SCRIBE_ROOTis writable; the server stores SQLite databases under the repo by default.
Contributing
Run
pytest MCP_SPINE/tests.Execute
python MCP_SPINE/scripts/test_mcp_server.py.Update docs (
MCP_SPINE/docs/dev_plans/scribe_mcp/…) and log your changes withMCP_SPINE/scripts/scribe.py.
Alright, here’s the no‑BS rundown for wiring your custom Scribe MCP server into Codex CLI on WSL. Follow the steps; swear at me later if it still doesn’t work.
1. Set up Scribe in your WSL project
Clone / extract the repo and install dependencies. From
~/projects/Scribe, create a venv and install Python deps:cd ~/projects/Scribe python -m venv .venv source .venv/bin/activate pip install -r requirements.txtScribe needs Python 3.11+; you already knew that.
Decide on a storage backend.
SQLite (default): do nothing – Scribe will drop a
.sqlitefile in your project directory.PostgreSQL: set
SCRIBE_STORAGE_BACKEND=postgresandSCRIBE_DB_URL=postgresql://user:pass@host/dbbefore launching the server.
Pick a root for your docs/logs. Scribe uses
SCRIBE_ROOTto find your project configs, dev plans and logs. Point this at theMCP_SPINEdirectory of your Scribe repo (e.g./home/austin/projects/MCP_SPINE/scribe_mcp). This is where the per‑project configs live underconfig/projects/– you can drop new YAML/JSON files there for each subproject.
2. Smoke‑test the server locally
Before hooking it to Codex, make sure the damn thing responds.
In another terminal you can run the provided smoke test:
The test fires up a short STDIO session, calls tools/list and checks that the server advertises tools. Fix any errors here before blaming Codex.
3. Tell Codex about your server
Codex CLI supports connecting to custom MCP servers via codex mcp add <name> ... or by editing the ~/.codex/config.toml file. The official docs confirm both methods.
Option A – use the CLI
Run this once to register Scribe as an MCP server (adjust paths/envs accordingly):
The
--envflags set environment variables for the server. You can repeat--envforSCRIBE_DB_URLif using Postgres.Everything after
--is the actual command used to launch the STDIO server.Codex stores this in
~/.codex/config.tomlfor you.
You can verify it worked with codex mcp ls or by launching codex, entering its TUI, and typing /mcp. Healthy servers will show up there.
Option B – edit the config file
If you prefer hand‑editing, open ~/.codex/config.toml and append the following:
Codex reads this file on startup; the docs note that each MCP server lives under its own [mcp_servers.<name>] table with command, args and env entries. Save it and relaunch codex.
4. Kick the tires
Launch Codex:
codexOnce inside the TUI, list MCP servers with
/mcp. You should seescribelisted as healthy.Ask Codex to call a tool, e.g.
/mcp call scribe tools/listto see what Scribe advertises. If that works, you’re good.
5. Dealing with multiple projects
Each project/subproject should have its own config file under
MCP_SPINE/config/projects/. Name the file after the project (e.g.myapp.jsonormyapp.yaml) and define settings likedefaults.reminderthere.Use Scribe’s
set_projecttool to switch between them at runtime; the server usesSCRIBE_ROOTto locate these configs.Scribe logs docs/changes under
docs/dev_plans/per project, so keep your directory structure tidy.
6. Run it under Windows
Codex CLI’s Windows support is experimental; OpenAI recommends WSL—you’re already using it. The commands above should work unchanged in your WSL shell. If you need to call it from a Windows terminal, use the \\wsl$\Ubuntu\home\austin\projects\Scribe style paths.
That’s it. Set the environment variables, smoke‑test the server, register it with Codex via codex mcp add or config.toml, and verify with /mcp in the Codex TUI. Once it’s hooked up, you can start logging and updating docs across your codebases. Go forth and build your Frankenstein of subprojects—Scribe will keep the mess slightly less messy.
This server cannot be installed
local-only server
The server can only run on the client's local machine because it depends on local resources.
Enables maintaining consistent project documentation and progress logs across development workflows. Provides tools for logging changes, tracking project phases, and keeping architecture docs synchronized with day-to-day development work.