Fellowship

Your AI Agents Remember Everything.

Persistent memory, auto-generated specs, and code review for coding agents.
Day 30 agents know what they learned on day 1.

$ npm install -g fellowship-cli

See what makes it different ↓

The problem with AI coding agents

Every tool gives you an agent.
None of them give it a brain.

Amnesia

Every run starts from zero. Your agent makes the same mistakes it made yesterday.

No plan

You say "add auth" and pray. No spec, no criteria, no way to verify the output.

No review

Code goes straight to your branch. No second pair of eyes. No quality gate.

What makes Fellowship different

Agents that learn.

Fellowship stores learnings, gotchas, and architecture decisions. Every run builds on the last.

Day 1

$ fellowship run "add API endpoint"

spec generated

writing code...

done

learned → "Use bcrypt rounds ≥ 12"

learned → "Validate JWT server-side"

memory: 2 learnings saved

Day 30

$ fellowship run "add payment system"

spec generated with 47 learnings

retrieved → "Always use decimal for money"

retrieved → "Validate webhook signatures"

retrieved → "Idempotency keys on all charges"

writing code (already knows your patterns)

zero mistakes. first try.

Same agent. Same codebase. 30 days of accumulated wisdom.

The full pipeline.

From planning to shipping. One CLI.

Spec-First Runs

Every task auto-generates a spec with acceptance criteria from your codebase context. The agent implements against the spec, not vibes.

$ fellowship run "add user roles" --review-spec

  spec: 002-user-roles.md 8 acceptance criteria 3 gotchas applied from memory
review spec before implementation? [Y/n]

Writer-Reviewer

One agent writes, another reviews against the spec. Automated quality gates with human override via --review.

 writer (sonnet) implementing...
  reviewer (opus) checking criteria...
8/8 criteria met
scribe updated CHANGELOG.md
ready to push

Interactive TUI

Full terminal dashboard. Browse run history, side-by-side diffs, learnings, decisions, and agent stats. Navigate with vim keys.

$ fellowship
Fellowship v0.4.8 > Run a task
Status
    Memory
    Doctor
    Search
jk navigate  Enter select  q quit

Decision Memory

Every architectural decision is captured, categorized, and stored in SQLite. When generating specs, only relevant decisions are retrieved via LLM filtering.

 categories: architecture backend frontend devops convention design bugfix other run #12 → "Use bcrypt rounds >= 12" run #15 → "Decimal for money, never float"

Role-Based Agents

6 built-in role templates with non-negotiable boundaries. Backend agents can't touch frontend. DevOps agents can't modify app logic. Hard constraints, not suggestions.

$ fellowship hire

  what role? backend engineer scope: APIs, databases, auth
must not: modify frontend components
must not: change CSS or UI layouts
owns: src/api, src/services, src/models

Auto Documentation

The Scribe agent updates your CHANGELOG, README, and API docs after every run. Documentation that writes itself.

$ fellowship run "add payments" --scribe
...
scribe: updated CHANGELOG.md
  scribe: updated docs/api.md
  scribe: updated README.md
3 docs updated automatically

Background + Notifications

Launch a task in background, close the terminal, go do something else. Get a Telegram message when it's ready for review. Slack and desktop notifications too.

$ fellowship run "add auth" --bg --review
Running in background (PID 41042) You'll be notified when it's done. ... 3 minutes later, on Telegram: Fellowship — review ready
Task: add auth Agent: forge Run: fellowship review

Simple commands. Powerful results.

Everything from setup to shipping.

fellowship --help

fellowship init Set up a new project from scratch

fellowship adopt Add Fellowship to an existing codebase

fellowship hire Build your agent team with role boundaries

fellowship run "task" Spec → write → review → done

fellowship run "task" --review Human-in-the-loop diff approval

fellowship spec "task" Generate spec without implementing

fellowship memory Browse runs, diffs, learnings, decisions

fellowship search "query" Full-text search across all project memory

fellowship status Interactive dashboard with agent stats

fellowship doctor Health check with 13-point diagnostic

fellowship scribe Auto-generate docs from current code

Stop prompting.
Start building.

One install. Zero config. Agents that actually remember.

$ npm install -g fellowship-cli

Works with

Claude Code

Codex

Ollama

Any CLI-based coding agent. Cloud or local.