Open source · MIT licensed

A coding agent,
harnessed end to end.

harnext reads, writes, and edits code, runs shell, and drives MCP — on open-source, local, or any provider. Run it interactively, hand it a task, or let it ship unattended.

harnext

The coding agent harness

Reads, writes, and edits code, runs shell, and drives MCP — on open-source, local, or any provider. The whole harness your team runs in the terminal.

~/projects/api — harnext

❯

Read the docs →

Open source

Built in the open.
MIT licensed, end to end.

harnext and Context Engine are free and open source. Read the code, open an issue, or ship a change — the whole harness was built to be forked.

QualityUnit / harnext Star

The coding agent harness — tools, providers, sandbox, and a runner that ships work unattended.

TypeScriptMIT

QualityUnit / context-engine Star

Token-efficient context for any harness — continuous indexing, ranking, and caching over a small HTTP + MCP surface.

TypeScriptMIT

Book a callPrefer a walkthrough? We'll show you harnext + Context Engine on your stack.

01 — The harness · harnext

From a one-line task to a reviewed change.
On rails.

harnext ships the agent, the harness, and the runner — no glue scripts, no agent server. Point it at open-source or custom models to keep development in-house, and hand your team a harness that works out of the box: it reads, writes, and edits files, runs shell, drives MCP servers, then runs the whole loop as a staged pipeline you can hand off and walk away from.

harnext --provider anthropic

❯ refactor packages/api/auth — split the session helper

read packages/api/auth/session.ts · 218 lines

grep "createSession" · 9 matches · 4 files

plan · 3 steps — extract tokens.ts, thin the helper, update imports

edit packages/api/auth/tokens.ts +47 −0

edit packages/api/auth/session.ts +8 −52

bash npm test -- auth ✓ 23 passing · 1.2s

↑ 3.1k ↓ 0.9k tokens · change ready for review

Interactive REPLA terminal agent that reads, writes, edits, runs bash, and drives MCP servers. One-shot mode with -p for scripts and automation.

Open & custom modelsAnthropic, OpenAI, Google, Ollama, NVIDIA + 20 more — or point --base-url at a local or self-hosted endpoint and keep everything in-house.

Drives any agentRun interactive yourself, or hand a stage to claude-code or codex with sandbox flags wired in.

Runs unattendedRegister a runner on your own hardware as a systemd / launchd service. Idempotent install; picks up where it left off.