Five Free Plugins That Stop Claude Code From Starting Over Every Session

Claude Code resets every session, burns tokens on the wrong model, and forgets your project. Five free plugins fix each of those problems. Setup takes under ten minutes.

You open Claude Code. You paste context from last week because it remembers nothing. You hit a usage limit mid-task and the session stops dead. You burn tokens on a model that is heavier than the task needed.

None of those are model problems. They are setup problems.

Five free plugins close each gap. One routes you to cheaper models automatically so your session keeps running. One gives Claude persistent memory so it never asks you to re-explain your project. One compresses your context window before it reaches the model, cutting token cost without touching your output quality. One scans your codebase and tells you exactly which hooks, skills, and MCP servers belong in your setup. One watches your working patterns and quietly improves your other tools in the background.

Together they change Claude or Codex from an assistant you constantly manage into a setup that manages itself.

If you have already installed the add-ons that give Claude Code hands, you have covered browser testing, database access, and design integration. Those add-ons connect Claude to your external tools. These five work differently. They change how Claude handles its own resources, memory, and self-improvement between sessions.

What "Plugin" Means Here

Two types of additions exist for Claude Code.

MCP servers are external processes you add from the terminal. Claude can call them as tools during a session. Install them with claude mcp add. They extend what Claude can do.

Native plugins install from inside Claude Code itself. They modify how Claude routes requests, stores context, and observes its own behavior. They are not tool calls. They change the runtime.

Three of the five plugins below are MCP servers. Two are native plugins. The resource guide linked at the bottom of this article has the exact install command for each one. Run each command, then type /mcp inside Claude Code. You should see the server show up as connected. That is your verification that the install worked.

One more thing before you start. Run claude --version. You need v2.1.0 or later for the native plugin system to work. Run node --version. You need Node 18 or higher. Every MCP server in this list runs on Node, and an old version is the most common reason a server shows up disconnected.

Both checks take ten seconds. If either fails, fix it before you continue.

OmniRoute: Your Session Never Stops

What it is. OmniRoute connects Claude Code to more than 200 free AI providers. It sits on top of your model routing layer and monitors your usage status across all of them.

What it does in practice. You are mid-session. A complex refactor. An automated pipeline halfway through. You hit the usage limit on your primary model. Without OmniRoute, the session stops. You copy your context, open a new window, and start again.

With OmniRoute, the moment that limit triggers, OmniRoute detects it and routes the next request to the best available alternative model. Your session continues without interruption. You do not see the wall. You do not lose your place.

This matters most if you run long sessions or fan out into parallel tasks. A usage limit on one model does not kill your work. OmniRoute moves you forward.

Install type. MCP server. Add it from your terminal and verify with /mcp once it is running.

The one thing that breaks it. The replacement model may be less capable than the one you were using. OmniRoute prioritizes continuity. It does not guarantee quality parity across every switch. If you get an answer that seems weaker or less precise after a session has been running a while, check which model OmniRoute is currently routing to. On a long or complex task, you may want to pause and reconnect to your preferred model before continuing.

Claude Code Setup: The Official Codebase Scan

What it is. This is an official Anthropic-built plugin. It does not come from a third-party developer. Anthropic ships and maintains it.

What it does in practice. Point it at your project. It reads your codebase structure, your existing configuration, and your installed tools. Then it outputs a concrete recommendation: which hooks make sense for how your project is organized, which Claude skills belong in your setup, which subagents would reduce manual prompting, and which MCP servers fit the tools your code actually calls.

It also runs the other direction. It flags the tools and configurations you have installed but are not using, and recommends removing them. A bloated setup costs tokens on every session. Claude Code Setup tells you what to cut.

Run it once when you start a project. Run it again after any major structural change. It reads your real codebase, not a generic template. Its recommendations reflect what is actually in your files.

Install type. Native plugin. Install from inside Claude Code rather than the terminal.

The one thing that breaks it. It reads what is there. If your project is early-stage with sparse structure, the scan returns sparse recommendations. Its value grows as your codebase does. Run it again once the project has substance.

Headroom: Fewer Tokens, Same Output

What it is. Headroom sits between you and the model. You do not interact with it. You do not see it working.

What it does in practice. Every prompt you send carries context: previous turns, instructions, file contents, thread history. That context accumulates. On a long session, a large portion of your token budget is context you already processed and do not need again.

Headroom intercepts the context before it reaches the model. It reads the full window, identifies what is load-bearing for your current prompt, and compresses or filters the rest. What arrives at the model is leaner. The output reflects the same intent. The token count is lower.

This is not summarization. Headroom does not paraphrase your previous instructions into something vaguer. It is a relevance filter. High-signal context passes through. Low-signal context, redundant passes, repeated boilerplate, gets compressed or dropped.

If you operate on a metered plan, token savings translate directly to cost savings. If you operate on a subscription, Headroom delays context-limit failures on long sessions and keeps each run lighter.

Install type. MCP server.

The one thing that breaks it. Headroom makes decisions about relevance. Most of the time those decisions are right. Occasionally they are not. On tasks where a detail from an early turn is the key to a late-turn problem, and Headroom filtered that detail as low-signal, you will get an answer that misses it. If output feels disconnected from context you know you provided, that is the first thing to check. You can always surface the relevant context explicitly in your next prompt.

Claude Mem: Memory Across Every Session

What it is. Claude Mem gives Claude Code persistent memory. When you close a session and return the next day, Claude knows who you are, what project you are working on, and what decisions you made.

What it does in practice. Today, every time you start a new Claude Code session, you start from nothing. You explain the project. You re-explain the file structure. You re-establish conventions and constraints that you defined two sessions ago.

With Claude Mem installed, that context lives between sessions. Your project description, your naming conventions, your architectural decisions, the files Claude has already read and understood, all of it persists. You open Claude Code and ask your question. You get an answer from something that already knows your setup.

Claude Mem stores structured memory tied to your project, not a vague log. You can review what it has stored. You can correct it. It is not a black box.

Install type. MCP server.

The one thing that breaks it. Memory is only as good as what it captured. If Claude stored a wrong assumption in an early session, it carries that assumption forward. Wrong memory is faster than no memory, which makes it worse than no memory. Review stored context after major project changes. After a pivot in how your codebase is structured, clearing stale memories and letting them rebuild is cleaner than correcting them one by one.

Task Observer: The Plugin That Improves the Others

What it is. Task Observer watches your Claude Code sessions as you work. You do not direct it. You do not see reports from it. It runs in the background.

What it does in practice. It logs what you do. Which prompts you run. Which tasks succeed without correction. Which ones you fix immediately. Where you send the same type of prompt repeatedly. It learns your working patterns from real usage data, not from instructions you give it.

Over time, it uses those observations to tune your other installed skills and workflows. If you run a particular kind of task ten times a week and always make the same correction on step three, Task Observer surfaces that pattern and adjusts how that skill runs. You start needing to make that correction less often.

It is less a tool you use and more a tool that uses your sessions to make everything else work better. The improvements are not dramatic from week one. They compound.

Install type. Native plugin.

The one thing that breaks it. Task Observer needs volume to work. Run Claude Code occasionally and it has little to observe. The feedback loop is thin. If you are using Claude Code for the first time or in short, infrequent bursts, install it now and let it run in the background. You will not see value in week one. By week six, the tuning is noticeable.

Which Ones to Install First

Do not install all five at once and then try to diagnose what broke. Install one. Verify it shows connected. Run a real task. Check the output. Then move to the next.

Here is the order that makes sense.

Start with Claude Code Setup. It is the diagnostic. It reads your actual codebase and tells you which of the other four your project needs most. If OmniRoute, Headroom, or Claude Mem appear in the output, you have confirmation from something that read your files. Install those next.

Install Claude Mem second. Memory payoff is immediate. From the first session after install, Claude stops asking you to re-explain. That alone changes the daily experience of using it.

Install Headroom third. Token compression applies to every session from that point forward. Once Claude Mem is working, your sessions carry structured context across days. Headroom keeps that context lean. Install Headroom after Mem and both run better together.

Install OmniRoute fourth. If you hit usage limits regularly or run long uninterrupted sessions, this is high value. If you rarely hit a cap, it is lower urgency. Still worth having for the sessions where it matters.

Install Task Observer last. It observes everything you have installed and improves it over time. Put it in after the other four are running. It will have more to observe from session one, and the feedback loop builds faster when the full setup is already active.

Honest Limits: What This Setup Does Not Fix

These five plugins change how Claude Code manages itself. They do not change the quality of your prompts, the structure of your tasks, or the architecture of your project.

OmniRoute auto-switches models, but a weaker model on a complex task produces weaker output. Headroom compresses context, but it does not improve bad instructions. Claude Mem stores what it captures, but a badly structured memory is a faster path to the wrong answer. Task Observer tunes your other tools, but it cannot compensate for a workflow that is fundamentally unclear.

Setup improvements compound good fundamentals. They do not replace them.

None of these plugins requires a paid subscription. All five are free to install. The resource guide linked at the bottom of this article has the exact install command for each one and the verification step to confirm the install worked.

One Thing to Do Today

Open Claude Code. Run claude --version. If you are on v2.1.0 or later, install Claude Code Setup first.

Let it scan your project. Read what it recommends. That output is the shortest path to knowing which of these five plugins will do the most for your specific setup.

If you only do one thing from this article: run the scan. Everything else follows from what it tells you.