whereismyai › Stacks
Claude Code setup for a solo developer
Nine layers around Claude Code, one pick each: current docs, a code map, memory, skills, parallel agents, tasks, costs and checks. What each costs, and the problem that says it is time to add it. From the setup that builds whereismyai.
For software engineers and founders. One person shipping a product with Claude Code, often with more than one agent at work. By whereismyai. Updated 18 October 2026, prices checked the same day.
Start with Claude Code on the $20 Pro plan, a one-page CLAUDE.md, plan mode and, from day one, checks the agent cannot skip. Add each other layer only when its problem shows up: Context7 for current docs, Graphify for a code map, MemMachine for memory shared between agents, skills for repeated instructions, one git worktree per agent with a handover file, Linear for tasks and ccusage for costs. Past the Claude plan, every layer is free, open source or has a free plan that covers one person.
The stack at a glance
Claude Code setup for a solo developer: each layer, its pick, when to add it and what it costs
| Layer | Pick | Add when | Cost |
|---|
| Base: The coding agent | Claude Code | Day one. | Pro at $20 a month covers evenings and weekends; Max, from $100 a month, a full day of agent work. |
|---|
| Context: Current docs, not remembered ones | Context7 | The agent calls a function that no longer exists, or you keep pasting documentation into the prompt. | Free for 1,000 calls a month, then 20 a day; Pro $10 a seat a month. Each call adds a few thousand tokens to the context. |
|---|
| Code map: Know the codebase without re-reading it | Graphify | The agent re-reads the same files every session, or the biggest files cost a slice of the context each time they are opened. | Open source. A shared deployment costs hosting and a token to look after. |
|---|
| Memory: Remember decisions between sessions and agents | MemMachine | You explain the same decision to the agent for the third time, or two agents disagree about a rule you settled last week. | Free; self-hosted on a small server or your own machine. |
|---|
| Workflows: Stop repeating the same instructions | Skills: one folder of instructions per job | You have typed the same instructions three times. | Free; an hour to write the first one well. |
|---|
| Parallel: Run several agents without collisions | One git worktree per agent, and a handover file | You start a second agent and it overwrites the first one's work, or you cannot say which agent owns which change. | Free; disk space and a few minutes per handover. |
|---|
| Tasks: Know what each agent is doing and what comes next | Linear | Two agents, five branches, and you have lost track of who is doing what. | Free up to 250 issues; Basic $10 a user a month, billed yearly. |
|---|
| See it: See what the agents did and what it cost | ccusage | You cannot say what last week cost, or whether the agents did more than you think. | Free (npx ccusage). |
|---|
| Verify: Catch mistakes before they ship | Gates the agent cannot skip | Always. Add a review pass once the changes outgrow what you read yourself. | Free; CodeRabbit reviews public repositories free. |
|---|
Base: The coding agent
The pick: Claude Code. Terminal-first, so it runs where the repo, the tests and the deploy tools already are, and it is what builds whereismyai: the worker, the nightly pipeline, this page. Two habits before any add-on: a CLAUDE.md of about a page (how to run the project, what never to touch) and plan mode for anything that spans more than a file.
Add when: Day one.
Cost: Pro at $20 a month covers evenings and weekends; Max, from $100 a month, a full day of agent work.
Already built in: CLAUDE.md, plan mode, slash commands, and the memory, subagents, hooks and worktrees the layers below lean on.
Alternatives: Cursor (The same work inside an editor. Plus is $8 a month with Grok models only; Premium is $30, with other makers' models billed by use.); Kilo Code (Open source extension, free with your own API key (you pay the provider); Kilo Pass credits from $19 a month.)
Context: Current docs, not remembered ones
The pick: Context7. Pulls the version of a library's docs you actually use into the context before the agent answers, which ends the code written against an API that changed last spring. An MCP server; the free plan is enough for one person.
Add when: The agent calls a function that no longer exists, or you keep pasting documentation into the prompt.
Cost: Free for 1,000 calls a month, then 20 a day; Pro $10 a seat a month. Each call adds a few thousand tokens to the context.
Already built in: Claude Code fetches pages and searches the web on its own. Context7 adds docs that are indexed and pinned to a version.
Alternatives: GitHub MCP server (Issues, pull requests and code search in the context.)
Code map: Know the codebase without re-reading it
The pick: Graphify. Builds a graph of the files, the symbols and the links between them, so the agent asks the graph where something lives instead of reading its way through a 6,000-line file every session. Ours runs as an MCP server that Claude Code and Cursor share. One honest line: on 10 October 2026 it rejected our token for a morning. Every remote layer is one more thing that can be down, so the skill that uses it says what to do without it (plain search).
Add when: The agent re-reads the same files every session, or the biggest files cost a slice of the context each time they are opened.
Cost: Open source. A shared deployment costs hosting and a token to look after.
Already built in: Claude Code's Explore subagent searches the repo on demand; it rebuilds its picture each time.
Memory: Remember decisions between sessions and agents
The pick: MemMachine. An open source memory layer (Apache 2.0) with an MCP server: what happened in a graph database, lasting facts about the project in SQL, shared by every agent you run. A decision made in Cursor on Monday is known to Claude Code on Tuesday. Self-hosted, so it is one more service to keep up: ours failed to connect the same morning Graphify did.
Add when: You explain the same decision to the agent for the third time, or two agents disagree about a rule you settled last week.
Cost: Free; self-hosted on a small server or your own machine.
Already built in: Claude Code's own memory: a folder of one-fact notes it writes and reads itself, with an index it loads each session. Enough for one agent; it does not reach Cursor or Kilo.
Alternatives: Mem0 (Hosted memory with a free tier, if you would rather not run a server.)
Workflows: Stop repeating the same instructions
The practice: Skills: one folder of instructions per job. A skill is a SKILL.md with the steps for one job, loaded only when that job comes up: how we check a claim before saying "done", how to build the code graph, how to run the app. It replaces the paragraph you were pasting into every prompt, and Cursor and Codex read the same format now. Hooks are the other half: a command that runs before or after a tool, such as lint after every edit, that the agent cannot forget.
Add when: You have typed the same instructions three times.
Cost: Free; an hour to write the first one well.
Already built in: Skills, slash commands and hooks are all part of Claude Code.
Alternatives: Anthropic's skills (Ready-made skills to start from.)
Parallel: Run several agents without collisions
The practice: One git worktree per agent, and a handover file. Each agent gets its own checkout of the repo on its own branch, so two never edit the same working tree: here, Claude Code on main, a second worktree for the advisor chat, Kilo's pool of worktrees for small jobs. The handover file is the other practice: when one agent hands to another, it writes the state, the rules and the open questions in one file the next agent reads first. HANDOVER.md is how Cursor handed this project back to Claude Code.
Add when: You start a second agent and it overwrites the first one's work, or you cannot say which agent owns which change.
Cost: Free; disk space and a few minutes per handover.
Already built in: Claude Code opens a worktree for a subagent on request and runs subagents inside one session. The handover file is yours to write.
Alternatives: Conductor (A Mac app that runs several Claude Code agents, each in its own worktree.)
Tasks: Know what each agent is doing and what comes next
The pick: Linear. An issue per job and an agent per issue, with Linear's MCP server so the agent reads the issue and marks it done. The free plan, unlimited members and 250 issues, is plenty for one person. We keep a small tracker of our own as well, milestones, tasks and open questions in the project's database, because we wanted them on the same page as the pipeline's numbers.
Add when: Two agents, five branches, and you have lost track of who is doing what.
Cost: Free up to 250 issues; Basic $10 a user a month, billed yearly.
Already built in: Claude Code keeps a task list inside a session. It does not outlive the session or reach another agent.
Alternatives: Vibe Kanban (Open source board that starts and tracks coding agents.); GitHub Issues (Free, next to the code.)
See it: See what the agents did and what it cost
The pick: ccusage. Reads Claude Code's local logs and prints tokens and cost by day, by session and by five-hour block, live if you want. Run it weekly: the sessions that cost most are usually exploration, not writing, which is the case for the code map and the memory above. For what the agents did rather than what they spent, we read the graph's own view and our tracker's pipeline page.
Add when: You cannot say what last week cost, or whether the agents did more than you think.
Cost: Free (npx ccusage).
Already built in: Claude Code's /cost shows the session you are in; ccusage adds the history.
Alternatives: Claude Code Usage Monitor (A live terminal dashboard of the current block.)
Verify: Catch mistakes before they ship
The practice: Gates the agent cannot skip. Three run here. Lint and type-check on every change (oxlint and tsc), scripts that exercise the search and the discovery pipeline before anything is called done, and a permission rule that refuses any write to the production database or a deploy from an agent: those commands are handed to a person to run. The last one matters most. An agent that cannot touch production cannot break it at two in the morning.
Add when: Always. Add a review pass once the changes outgrow what you read yourself.
Cost: Free; CodeRabbit reviews public repositories free.
Already built in: Claude Code: /code-review on a branch, hooks to run lint after every edit, and permission rules per command.
Alternatives: CodeRabbit (AI review on every pull request; free on public repositories.); Sentry Seer (Error tracking that proposes the fix, once there are users.)
Questions
What is the best Claude Code setup for a solo developer?
Start with Claude Code, a CLAUDE.md of about a page (how to run the project, what never to touch), plan mode for anything that spans more than one file, and checks the agent cannot skip. Then add a layer only when its problem appears: Context7 for current docs, Graphify for a code map, MemMachine for shared memory, skills, worktrees, Linear and ccusage. Prefer a different base agent? See the Claude Code alternatives.
How many Claude Code agents should I run in parallel?
Start with one, and add a second when you have separate work to give it. Give each agent its own git worktree on its own branch so two never edit the same working tree, and keep a handover file with the state, the rules and the open questions for the next agent to read first. Once you have two agents, several branches and no clear picture of who is doing what, give each agent an issue in Linear.
Which MCP servers should I start with in Claude Code?
Context7 first. It pulls the version of a library's docs you actually use into the context, so the agent stops calling functions that no longer exist, and its free plan is enough for one person. Add Graphify when the agent keeps re-reading the same files, MemMachine when you keep explaining the same decisions, and Linear's MCP server once you track tasks across agents. Each remote server is one more thing that can be down.
How much does a Claude Code setup cost per month?
From $20 a month, and nearly all of it is the Claude plan: Pro at $20 covers evenings and weekends, Max from $100 covers a full day of agent work. Context7 and Linear have free plans that cover one person, CodeRabbit reviews public repositories free, Graphify and MemMachine are open source, and ccusage is free. Self-hosting costs a small server. Run ccusage weekly to see what you actually spend.
How do I keep context and memory across Claude Code sessions?
Start with what is built in: a CLAUDE.md for the project's rules, and Claude Code's own memory, a folder of one-fact notes with an index it loads each session. That is enough for one agent, but it does not reach Cursor or Kilo. To share decisions across agents, add MemMachine, an open source memory layer you host yourself, or Mem0 for hosted memory with a free tier. When one agent hands work to another, write a handover file.
Changelog
- 18 October 2026: Added a short summary and five questions. Prices re-checked: Cursor's plans changed (Plus $8 with Grok models only, Premium $30), Max is from $100, Linear's Basic is billed yearly and CodeRabbit is free on public repositories.
- 10 October 2026: First version, written from the setup that builds whereismyai. Prices checked the same day.
Sources
- Claude plans and pricing, Anthropic
- Claude Code: skills, Anthropic
- Cursor pricing, Cursor
- Kilo Code pricing, Kilo
- Context7 plans, Upstash
- graphify, GitHub
- MemMachine, GitHub
- Linear pricing, Linear
- ccusage, GitHub
- CodeRabbit pricing, CodeRabbit