No tokens, no limits
Runs on your own machine with a local Ollama model. No API keys, no usage caps, no subscriptions.
Agentic Engine for Minimal Resources
Build whole projects with a local AI. No tokens. No limits.
Agemir is a command-line program written in C that turns local AI models into autonomous coding agents. It splits big projects into small tasks and runs each one in a short-lived Ollama session, even on low-end hardware.
- 0 API tokens
- Runs on your machine
- MIT license
- 5 interface languages
$ ./agemir Agemir - Agentic Engine for Minimal Resources 1. Run Agent on Existing Project 2. Generate New Project Template 3. Run Hardware Benchmark 4. Import Filled Project Template 5. Generate Project From a Description 6. Settings 0. Exit > 5
Local models have two big problems. Agemir fixes both.
Local models need a lot of RAM, and their context degrades in long chats. Agemir keeps every session small and keeps all state in plain files.
ProblemContext degrades in long chats
Agemir
Every task runs in a fresh, short-lived model session.
ProblemHigh RAM usage
Agemir
The model is unloaded after each task with ollama stop.
ProblemLost progress
Agemir
All state lives in plain files you can read and edit.
Everything you need to build with a local model
Describe a project once. Agemir builds it task by task, without human help.
Fresh context every task
Each task runs in a new, short-lived model session, so the model never gets lost in a long chat.
Gentle on RAM
The model is unloaded after every task, and a built-in benchmark picks one that fits your machine.
State in plain files
Progress, history and per-attempt logs are readable text files you can inspect and edit.
Pause and resume
Press Ctrl+C or drop a PAUSE file in the project folder, then continue exactly where you stopped.
Verified step by step
Each task can run a VERIFY command. Errors go back to the model, with up to 5 attempts per task.
Describe it in plain words
Let the agent write the project template for you. Agemir checks it with the real parser before starting.
Graphical projects
Built-in UI guidelines (navbar, login menu, hamburger menu, cards grid) and free icon and image links guide the design.
Speaks 5 languages
The interface comes in English, Spanish, French, German and Japanese.
How it works
From idea to working project in four steps.
Describe
Fill in a template, or describe the project in plain words and let the agent write it.
Split
The project becomes goals and small tasks. Each task writes one file or runs one command.
Build and verify
Every task gets a fresh model session. Verification errors are fed back, up to 5 attempts.
Resume anytime
If a task keeps failing, the run stops safely. Fix it and continue from that exact task.
One template. Many small tasks.
A project is a plain text template: a description, optional style rules, and a list of small tasks. Each task names the file to write and, if you want, a command that must pass.
PROJECT_NAME: Notes CLI
OBJECTIVE: A tiny command-line notes app
TECH_STACK: Python 3
GOAL: Core logic
TASK: Write the note storage module
FILE: notes/storage.py
VERIFY: python3 -c "import notes.storage"
TASK: Add the command-line interface
FILE: notes/cli.py
READS: notes/storage.py
VERIFY: python3 notes/cli.py --helpSimplified example. The repository ships a blank and a complete template in resources/templates.
Built for modest hardware
- A hardware benchmark picks a model that fits your RAM and speed.
- The model list marks each option as [OK] or [X] for your RAM and shows an estimated speed.
- Choose Auto, pick any listed model, or type your own Ollama tag.
- If memory runs out, Agemir stops at once and suggests a smaller model.
TipOllama's default context window is small. If long tasks look truncated, raise it with OLLAMA_CONTEXT_LENGTH or num_ctx.
Quick start
You need a C compiler (gcc). Agemir offers to install Ollama on first run.
1
Get the code
git clone https://github.com/wizt-jaegger/agemir.git
cd Agemir2
Build it
gcc -O3 -DOLLAMA_ENABLED=1 Agemir.c -o agemir3
Run it
./agemirChoose a language, enter your name, let the benchmark pick a model, then create a project with option 5 (from a description) or options 2 and 4 (by hand). Run it with option 1.
Want to explore the menus without calling a model? Build without the flag: gcc -O3 Agemir.c -o agemir. It runs in simulation mode.
Frequently asked questions
Does Agemir need internet access?
Only to install Ollama and to download a model the first time. Tasks then run on your own machine with a local model, so there are no API tokens or usage limits.
Which models can I use?
Any model available in Ollama. Agemir ships a catalog in resources/models.txt, marks each model as [OK] or [X] for your RAM, and lets you choose Auto or type your own tag.
How much RAM do I need?
It depends on the model. The built-in benchmark picks one that fits your machine, and smaller models run on modest hardware. If memory runs out, Agemir stops and suggests a smaller model.
What happens when a task fails?
Agemir retries up to 5 times, feeding the error back to the model and logging every attempt. If it still fails, the run stops, the file is restored and you are told where the logs are. Fix the problem and run again to resume from that task.
Can I pause a long run?
Yes. Press Ctrl+C or create an empty file named PAUSE in the project folder. Progress is saved, and the next run continues from the saved task.
Is it safe to let it run commands?
Tasks marked RUN execute shell commands, so review any template you did not write. Agemir rejects unsafe paths on import, and you should never store sudo passwords in a template.
Which languages does the interface support?
English, Spanish, French, German and Japanese.
Is Agemir free and open source?
Yes. Agemir is released under the MIT License, so you can use, modify and share it freely. The source code is on GitHub.