Behavior powered by Agentic SWE ↗

The Italian
Open Source
Coding Agent.

interference reads your code, edits files and runs commands through an agentic loop. Every turn passes through the open-source Agentic SWE behavior framework, with code-enforced permissions and a read-only Plan mode.

bun install -g interference-agent

Requires Bun 1.3+ — the runtime interference runs on.

GitHub ↗ Join Discord ↗

interference

/ 01 — open behavior framework

Prompts can suggest. The framework decides what happens next.

Every primary turn is evaluated by Agentic SWE — the separate open-source protocol and runtime at the heart of interference. It classifies intent, opens setup and planning gates, scopes capabilities, and requires host evidence before completion.

Explore Agentic SWE on GitHub ↗ Protocol 1.1 · MIT · separate repository

setup → plan → execute

Gates before action

Non-trivial work cannot skip repository setup or planning. Complex questions remain read-only when no mutation was requested.

protocol ∩ mode ∩ tools ∩ policy

Deny wins

The model may describe intent, but it cannot grant access. Agentic SWE and Interference can only narrow the toolset.

event → evidence → completion

Evidence before done

Only successful redacted host events satisfy hard criteria. A failed test never becomes proof of validation.

Agentic SWEowns behavior: classify · gate · route · verify

interferenceowns execution: models · tools · permissions · sessions · TUI

AGENTS.md → gates → scoped tools → evidence — behavior that survives the chat

/ 02 — the loop

An agent loop you can watch and trust.

01
You ask

Describe a task in plain language. interference plans before it touches anything.
02
It calls tools

read · ls · glob · grep · webfetch to explore, write · edit · bash · task · todowrite · question to act — multi-step, streamed live.
03
You stay in control

Every mutating action is gated by allow / ask / deny. Approve with a diff in front of you.
04
Undo anything

Sessions are persisted and snapshotted. Roll back a change the agent made, instantly.

/ 03 — skills

Teach it once. It remembers every time.

without skills

Generic agent

Every session starts from zero. You re-explain your commit style, your test setup, your release process — every time, in every conversation.

Re-explained every session
Generic, one-size-fits-all output
Inconsistent results

with skills

Scoped skills

Skills aren't one block of text bolted onto every prompt. interference injects each one exactly where it belongs — into the main conversation, or into a dedicated subagent with its own system prompt built around that one task.

Auto-detected, or called directly: /commit
Injected into the exact prompt that needs it
Subagents get their own tailored prompt too

mode · read-only

Plan

The agent explores and reasons — never writes. Perfect for understanding a codebase or scoping a change before committing to it.

Filesystem read & search only
Cites file:line evidence
Zero side effects

mode · full access

Build

The agent writes files, applies atomic edits and runs commands — each behind your permission rules, with destructive actions denied by default.

Atomic edit (unique-match)
Sandboxed paths, timed bash
Approve-with-preview

/ 04 — what's inside

Small surface. Serious capability.

⌘

Native tool-calling

A typed tool registry over the Vercel AI SDK — multi-step, with tool errors fed back for self-correction.

⛨

Permissions by design

allow / ask / deny rules enforced in code, not in the prompt. The model can't talk its way past them.

◫

A real TUI

Streaming output, spinners and live tool steps in a flicker-free terminal interface built with Ink.

↺

Sessions & undo

Resume where you left off. Snapshots before every mutation mean undo/redo on the agent's edits.

⌥

Cost transparency

See exactly what you’re spending on API calls — live, per-request, in your terminal. No hidden usage, no surprise bills.

⚡

Bun-fast

Runtime, test runner, and package manager — one toolchain, zero config. Starts instantly.

✦

Extensible skill system

Agent Skills format with auto-detection by keyword matching. Invoke via /skill-name. 3 skills bundled, your own are just a SKILL.md away.

⎔

Subagents

Delegate complex multi-step tasks to isolated agents — Explore (read-only), General (full access), Review (bug/security/simplicity findings), or your own custom agents. Launch several in the same turn to run them in parallel.

⌨

Slash commands + autocomplete

/help · /clear · /undo · /redo · /init · /model · /plan · /build · /compact · /sessions · /rename · /provider · /thinking. Fuzzy autocomplete on /.

/ 05 — why interference

Not just another agent.

European, by choice

Built in Italy. Open-source, MIT licensed, not VC-funded. Your code never leaves your machine unless you explicitly send it to an API. An independent, GDPR-native alternative to Silicon Valley agents.

Radically transparent

Every tool call, every reasoning step, and exactly what you’re spending on API costs — live, in your terminal. No hidden usage. No surprise bills.

Privacy by architecture

There’s no interference server. Chats and sessions stay local in ~/.interference/, never uploaded. We don’t save your conversations, don’t collect your data, and don’t train on anything — what you send your LLM provider is between you and them.

/ 06 — quickstart

In your terminal in a minute.

1 — Install Bun 1.3+ (skip if you already have it):

curl -fsSL https://bun.sh/install | bash

2 — Install interference: