FinishIt

Finally finish the project you started six months ago.

npx @finishit/finishit github.com/you/that-repo-from-2024

Website: nexus-prime.cfd/finishit

Point it at any abandoned repo. An AI agent indexes the codebase, identifies what you were stuck on, and ships a PR.

How it works

1. Indexes your repo — git history, file tree, README, open issues, every wip: commit and TODO comment land in a local SQLite at ~/.finishit/repos/<hash>/memory.db.
2. Diagnoses the stuck point. The first thing it tells you is what you were actually working on, quoting the real wip commit and the real TODO. No hallucinated paths — a guard rejects any output that names a file or hash not in the indexed facts.
3. Ships a PR in a git worktree on a finishit/<slug>-<ts> branch. Tests run if a test command is detected. Default branch is never touched. One PR per command. No autonomous loops.

Setup

# Option 1: use a CLI you already have (zero config — recommended)
brew install codex                       # uses your ChatGPT subscription
# or: npm i -g @anthropic-ai/claude-code  # uses your Claude subscription

# Option 2: BYO API key
export ANTHROPIC_API_KEY=sk-ant-...
export OPENAI_API_KEY=sk-...

# Option 3: local Ollama
ollama serve

Provider auto-detection order (first match wins):

1. claude CLI on PATH (--model claude-code to force)
2. codex CLI on PATH (--model codex to force)
3. ANTHROPIC_API_KEY
4. OPENAI_API_KEY
5. A reachable Ollama server at OLLAMA_BASE_URL or http://localhost:11434

Codex CLI (uses your ChatGPT subscription)

brew install codex
npx @finishit/finishit .
# or force it: npx @finishit/finishit . --model codex

Set FINISHIT_CODEX_MODEL to override the model sent to codex (default: gpt-5).

Claude Code CLI (uses your Claude subscription)

npm i -g @anthropic-ai/claude-code
npx @finishit/finishit .
# or force it: npx @finishit/finishit . --model claude-code

Ollama, no key

ollama serve
ollama pull qwen2.5-coder:14b
npx @finishit/finishit . --model qwen2.5-coder:14b

Use OLLAMA_BASE_URL when Ollama is not on the default port:

export OLLAMA_BASE_URL=http://localhost:11434
npx @finishit/finishit . --model qwen2.5-coder:14b

Anthropic

export ANTHROPIC_API_KEY=sk-ant-...
npx @finishit/finishit . --model claude-sonnet-4-6

OpenAI

export OPENAI_API_KEY=sk-...
npx @finishit/finishit . --model gpt-5

Passing --model forces routing: codex → Codex CLI, claude-code → Claude Code CLI, claude-* → Anthropic API, gpt-* and supported o-series prefixes → OpenAI API, anything else → Ollama. No account, no signup, no telemetry.

What it does NOT do

• No web UI, no dashboard, no IDE plugin.
• No team features, no shared memory, no Slack/Linear integration.
• No code leaves your machine except to your configured LLM provider.
• No autonomous loops or scheduled runs.

Built on Nexus Prime

FinishIt is the smallest possible surface of the Nexus Prime engine — Memory Fabric, Ghost Pass, Session DNA, and worktree-isolated execution. If you want this for your team's monorepo with shared memory across every agent your team uses, look at nexus-prime.cfd/teams.

MIT.