Spec-driven orchestration for AI coding agents

Turn AI coding agents into a spec-driven delivery pipeline

Conflux orchestrates Claude Code, Codex, and OpenCode through proposals, parallel worktrees, acceptance gates, and merge-ready changes.

Spec-firstParallel worktreesAcceptance gate

Try it in 5 minutes View GitHub

Install Conflux

$ cargo install cflx

Conflux Run

change/add-search-filter

72%applying

change/fix-auth-flow

acceptance pass

change/refactor-cache

queued

proposal→apply→acceptance→archive

Trusted primitives for agentic delivery

Conflux keeps AI implementation work inside units that developers can inspect, run, accept, and merge.

Spec-first changes

Fix proposal and tasks before implementation so every AI work unit has a clear boundary.

Parallel worktrees

Run each change in an isolated git worktree and move multiple implementations safely.

Acceptance gate

Separate implementer and evaluator, then move only passed changes toward archive and merge.

One-shot agent coding breaks down at product scale

Single prompts are fast, but real products need visible state, isolated changes, and a reliable done signal.

Spec and intent drift inside long chat threads

Changes collide in one working tree

Implementation and review blend together

Humans keep polling for whether the work is actually done

How Conflux works

A delivery pipeline for AI coding agents: define the change, run it in isolation, evaluate it, then merge the result.

Proposal

Freeze intent and scope into a change unit.

Parallel Worktrees

Advance each change in an isolated git worktree.

Apply Loop

Let implementation agents iterate inside small context.

Acceptance Gate

Separate implementation from pass/fail judgment.

Merge-ready

Archive, resolve, or merge the changes that pass.

Why not just one CLI agent?

Conflux is the control plane around the agents you already use.

Standalone CLI agent

Conflux

Starts from one prompt

Starts from a spec / proposal

Work unit can become vague

Tracks every change unit

Implementation and review often mix

Separates apply and acceptance

Depends on one long context

Runs isolated parallel worktrees

Done judgment returns to the human

Flows toward archive / resolve / merge

Built for teams that treat AI coding as delivery work

Not a toy prompt wrapper. Conflux fits teams that want AI agents inside an auditable engineering workflow.

Built for

Teams already using Claude Code, Codex, or OpenCode
Developers who do not want one-shot prompt coding to be the process
Teams separating spec, implementation, review, and merge
Projects running multiple safe changes in parallel
Humans who want to focus on intent and final judgment

Not for

Generating a single throwaway file
Light autocomplete as the whole workflow
Avoiding git, spec, or review flow management

Start in 3 commands

Install Conflux, initialize a project, and start the local control loop.

Install

cargo install cflx

Init

cflx init

Run

cflx

Need remote monitoring?

Use server mode with the Web UI dashboard, WebSocket updates, and REST API.

$ cflx server

dashboard: http://localhost:39876

ws://localhost:39876/ws

/api/workflows

Watch the demo

Try it in 5 minutes

FAQ

Short answers for teams evaluating Conflux.

Is Conflux itself a coding agent?+

No. Conflux is the orchestrator. It launches and coordinates external coding agents such as Claude Code, Codex, or OpenCode.

When is it a better fit than using one CLI agent directly?+

Use Conflux when you need persistent change tracking, parallel execution, review separation, or a workflow that continues after the original prompt ends.

Can I monitor runs remotely?+

Yes. Server mode and the Web UI expose a dashboard, WebSocket state updates, and REST endpoints for monitoring and automation.

Move from one prompt to a continuous AI development flow

Start with a local run, then grow into server mode when the workflow needs visibility.

Try it in 5 minutes View GitHub