Mycel Agent

v0.1.3 · open beta · local-first

The living network
of intelligence.

An autonomous AI agent that runs entirely on your machine — growing through your filesystem, remembering across sessions, never leaving your device.

GitHub How it works

Architecture

A loop that never stops growing

Every prompt enters a tight reasoning cycle. The agent thinks, acts through tools — its hyphae — observes the result, and repeats until the task is resolved.

01

Think

The model reads the prompt, working directory, and all prior tool results. It decides the next move.

02

Act

A hypha fires: bash, read_file, write_file, web_fetch, or memory_save.

03

Observe

Results return into context. The model reflects, adapts, and either continues or delivers its answer.

04

Remember

Key facts persist in a local SQLite database — recalled automatically in future sessions.

Quick start

Five steps to first run

Requires Node.js ≥ 20 and Ollama. No API keys. No cloud.

terminal 
# 1. Clone
$ git clone https://github.com/mycelagent/mycel.git && cd mycel

# 2. Install
$ npm install

# 3. Pull a model
$ ollama pull llama3.2

# 4. Start Ollama
$ ollama serve

# 5. Run
$ npx tsx src/index.ts task "summarize all TODOs in this codebase"

  ✓ ready
  ⟳ hypha: list_dir · read_file · read_file …
  ✦ Mycel completed

Commands

The full CLI

task

One-shot autonomous task. Give a prompt; the agent works until complete.
npx tsx src/index.ts task "add JSDoc to every function in src/"

chat

Interactive multi-turn session. Type exit to quit, memory to inspect memories.
npx tsx src/index.ts chat

demo

Seeds a neutral temp workspace and runs 4 tasks automatically — for screen recordings.
npx tsx src/index.ts demo

doctor

Checks Node version, Ollama status, model availability and loaded config.
npx tsx src/index.ts doctor

memory

List or delete persistent memories stored across sessions.
npx tsx src/index.ts memory list

sessions

Browse past sessions; resume any with --session <id>.
npx tsx src/index.ts sessions

Advantages

Why Mycel is different

Most AI tooling sends your data to a remote server. Mycel is built on the opposite principle.

Fully local

Runs on Ollama. Your code, prompts, and memories never leave your device. No API keys, no billing, no data policy to accept.

Persistent memory

SQLite-backed memory survives restarts. The agent recalls facts from previous sessions and grows with you over time.

Real action

Mycel doesn't suggest — it reads files, writes code, runs shell commands, fetches URLs, and saves knowledge. It acts.

Model-agnostic

Point mycel.config.json at any Ollama model. Swap without changing your workflow.

Minimal & hackable

Pure TypeScript. No runtime magic. Add tools, change the loop, fork the entire thing — the source invites modification.

CLI-native

No Electron, no browser tab. Runs anywhere Node runs. Pipe it into scripts, CI pipelines, or your own tools.

Market position

How it compares

Feature Mycel Cloud agents Local LLM UIs
Fully local execution yes no yes
Persistent cross-session memory yes yes no
Real tool use (files, bash, web) yes yes no
Zero subscription / no API key yes no yes
CLI-native, scriptable yes no no
Open source, fully hackable yes no partial