Foremerge

Why STEADY

STEADY (80) for a coherent, genuinely agent-native protocol with an excellent trust posture — Apache-2.0, local-first with no cloud dependency, a deterministic (no-LLM) detection path, an honest limitations section, and real traction (512 GitHub stars) for a young project. Held below VITAL because this is a public-surface review (medium confidence, no hands-on multi-agent run by Hlido), the project is explicitly pre-1.0 (0.5.0), multi-machine coordination is not yet in the box, and the authors have published no reproducible coordinated-vs-uncoordinated benchmark yet.

Public-surface checklist

• PASS Homepage loads (required) — foremerge.com returned a full product page (page_title 'Foremerge: catch intent conflicts before code conflicts')
• PASS Primary value prop (required) — 'Catch intent conflicts before code conflicts' — an open-source coordination protocol for coding agents, built above Git
• PASS Cta present (required) — 'Quickstart' / 'View on GitHub' / 'curl -fsSL https://foremerge.com/install.sh | sh'
• PASS Pricing or access — Free and open source (Apache-2.0); binaries for macOS/Linux, or 'cargo install --locked foremerge'
• PASS Evidence or demo — Worked quickstart with a two-agent example, a JSON conflict finding (destructive_vs_additive, rule FM-C001), and an intent→committed lifecycle diagram; benchmark fixtures cited but no published results yet

What it does well

• Names a real, emerging failure mode precisely — two agents making individually-correct edits in separate worktrees that Git merges without conflict but that undo each other's work
• Deterministic, explainable conflict detection with no LLM in the detection path — the same inputs always produce the same finding, which is the correct property for trusted infrastructure
• Maximally agent-native distribution: one Rust binary exposing a CLI, an authenticated JSON/OpenAPI loopback daemon, and an 18-tool MCP server over stdio, with a one-line 'tell your agent to install it' path
• Provider-agnostic by design — mixed Claude + Codex fleets are the design center, and every ChangeSet records which agent/model did what against an exact Git fingerprint
• Unusually honest public surface — an explicit 'what it does not claim' section, docs/limitations.md, advisory-not-locking coordination, and 'no performance claims until there are'

What it fails at

• Explicitly pre-1.0 (0.5.0) MVP with schemas that 'may still change' — an early dependency risk the authors flag themselves
• No published, reproducible coordinated-vs-uncoordinated benchmark yet, so the core value ('catch conflicts before they land') is argued rather than measured
• Local-first only — a single SQLite store shared across worktrees on one machine; multi-machine shared mode is on the roadmap, not shipped
• Deterministic heuristics can miss synonymous concepts and warn on compatible work, as the docs acknowledge — coverage is advisory, not guaranteed
• No hands-on run by Hlido — the conflict-detection quality, tool reliability, and lifecycle gating are asserted by the project, not independently confirmed here

Pricing & access

• ModelOpen source
• Free entry pointYes — a free tier or open-source edition is documented
• Pricing findable on the public surfacePASS Free and open source (Apache-2.0); binaries for macOS/Linux, or 'cargo install --locked foremerge' (tested 2026-09-28)

Derived from Hlido-held evidence only (engine checklist + editorial text); quotes are verbatim from the scorecard; not vendor-supplied; re-derived daily. Verify current prices on the vendor's pricing page. Last verified 2026-09-28.

Related agents

• Claude Code Coding · 91
• Aider Coding · 90
• @aliasjeff/acli Coding · 90

Agent relevance

API CLI MCP Behavioral-testable

Native, MCP-first: an agent (or its harness) installs the Foremerge binary, runs 'foremerge init' and 'foremerge setup all' per repo, then drives coordination through an 18-tool MCP server over stdio (publish_intent, claim_work, check_conflicts, publish_changeset, run_verification, accept_changeset, and more). A CLI with --json and an authenticated JSON/OpenAPI loopback daemon offer the same operations for harnesses preferring HTTP. Provider-agnostic by design — any MCP-capable client drives the identical tools.

Agent-friendly score: 9/10

Evidence

• Open-source (Apache-2.0) coordination protocol for coding agents built above Git — agents publish intent and semantic scopes before editing so conflicts surface while both worktrees are still clean — source (2026-09-28) verified
• Deterministic conflict detection with no LLM in the detection path — the same inputs always produce the same answer; coordination is advisory and never locks a file — source (2026-09-28) verified
• Ships as one binary exposing a CLI, a JSON/OpenAPI loopback daemon, and an 18-tool MCP server over stdio; coordination state lives in a SQLite store under the repo's Git common directory — source (2026-09-28) verified
• Provider-agnostic — mixed Claude + Codex fleets are the design center; every ChangeSet records which agent and model did what against an exact Git fingerprint — source (2026-09-28) verified
• Pre-1.0 (0.5.0) local-first MVP with no published coordinated-vs-uncoordinated benchmark yet and multi-machine coordination out of scope — stated by the project itself — source (2026-09-28) verified