Software development
Developer notes and project planning without losing the thread
A codebase has context outside the editor: the trade-off you rejected, the incident detail that matters next quarter, the sketch of an API, and the task that should exist before the next build. Give that thinking a durable home.
Developers already have a code editor, an issue tracker, and a terminal. What often goes missing is the developer notes and project planning layer between an idea and a merged change: architecture notes, decision records, debugging context, design questions, and a plan that still makes sense after a week away.
Mind Marshal is a local-first workspace for that layer. It gives you a rich block editor, formatted code blocks, real to-do lists, pages that link to each other, and databases for projects or release work. The workspace stays in one local SQLite file and supports Markdown import and export.
Use one page for one technical question
Start a page with the decision you need to make, the constraints, examples, and the smallest next step. Keep a code snippet next to the explanation it needs. Link related pages with [[mentions]] and use backlinks to recover the context around an API, an incident, or a product decision.
Turn an idea into an executable plan
- Frame the outcome. Write what will be true when the work is complete, not only the implementation you first imagined.
- Record open questions. Keep unknowns visible instead of hiding them in a comment or chat thread.
- Break down the next move. Use a checklist for the few steps that unblock research, implementation, review, or rollout.
- Track the work at the right altitude. Use a project database for active initiatives, with a board or timeline when that makes status or sequence easier to see.
Bring in selected repository context, not another noisy feed
The developer journal field guide includes a debugging log, a lightweight architecture decision record, and a review rhythm for promoting durable knowledge back into the repository.
If you connect GitHub with one-click sign-in or a fine-grained token, selected repository activity can be captured into your Inbox for personal triage. Capture-in remains a deliberate read path for context. When you explicitly press Push code in the editor, Mind Marshal can commit one file through the Contents API; it is still not a replacement for GitHub’s issue and review workflows.
Use AI as a reviewable thinking aid
Optional assistance can help turn a rough plan into tasks, research a selected topic, or query your workspace with source citations. It is never a reason to skip technical judgment. You choose whether to connect an AI provider, use your own API key, or keep model work local with Ollama.
What it does not replace
Mind Marshal is not an IDE, debugger, CI system, source-control client, or hosted issue tracker. Its role is to hold the reasoning and project context around those tools, then help you return to the right next step with less reconstruction.
How Pro takes this further
Let the workspace do the preparation, and keep every decision
The notebook is free. Pro is for the paperwork around a change: release stages that gate on evidence, decision records appended automatically, and a run that turns an incident thread into a postmortem draft you then own.
- Stage gates on a release board. A Ready-to-merge stage can ask the agent to check the card’s page for a test plan and rollback note and answer PASS or BLOCK. BLOCK routes to Needs work; PASS holds for your approval.
- ADRs that append themselves. An append to page action writes a dated line to your architecture-decision log whenever a design card moves to Decided, with the agent’s summary of the trade-off.
- Postmortems as a run. Scope a run to the incident page and the linked chat captures: “draft a postmortem with a timeline and a flowchart of the failure path”. The output is native blocks, undoable as a whole, and exports to Markdown for the repo.
- MCP for the agents you already use. Mind Marshal runs as a local MCP server. Claude Code or Codex can search and read the pages you grant, and can propose a page or an append that appears in your run rail with Accept and Reject. No external agent writes directly.
- Local models welcome. Every one of these works with Ollama on your own machine; the recorded fixtures behind them were exercised against qwen2.5:7b, not only hosted models.
A concrete run: Push a design card to Decided. The stage’s instruction asks for a one-paragraph trade-off summary; the action appends it under today’s date in ADR log and sets Status to Recorded. Later, Claude Code proposes an addendum through MCP; you accept it from the rail, and Undo is one click if it was wrong.
Everything above is optional and reviewable. Runs, stage workflows and chat use the model you connect (a ChatGPT, Claude or Kimi plan you already pay for, your own API key, or a local Ollama model), every change lands in Activity with Undo, and Free Core stays the unlimited manual workspace whether or not you ever turn Pro on. See how the AI wiring works.
Try the workflow with your own work
Start a local notebook for the work around the code.
Use Free Core to organize one technical initiative with its notes, code snippets, decisions, and next steps before you migrate anything bigger. The optional 14-day Pro evaluation covers AI and automation.