Local-first harness, not another model

Keep your AI CLIs. Lose the re-brief.

VibeFlow connects your existing agents, carries context between them, owns each CLI run, and requires proof before it says done.

Install VibeFlow Read the source
VibeFlow conversation Home and orchestration preview
npx @magicpro97/vibeflow

See a real run go from prompt to evidence

Follow a Codex run through dispatch, edits, verification, and recorded evidence in one local workflow.

Pick up, do not restart

Open any session and continue with its participants, queue, trace, and settings already in place.

Type while agents work

New messages wait in order, survive restarts, and stay editable until dispatch claims them.

Ask for tools in plain language

Capability requests move from chat to exact targets, review, approval, and an audited result.

Proof is part of the answer

Tests, evidence, scope, and confidence land before VibeFlow lets a run call itself done.

Where multi-CLI work usually breaks

VibeFlow keeps sends ordered, handoffs explicit, installs reviewed, and process ownership proved.

Resume, steer, inspect

One Home, no dashboard maze

Open a session and keep talking. Agents, capabilities, settings, and trace stay beside the conversation, not behind another control panel. Conversations auto-classify into per-project folders: import your repos, fast local FTS matching, AI only when it matters.

FIFO for your messages, editable before dispatch

Send now, queue safely

Keep typing while an agent works. Your messages stay ordered, failures stay attached to their payload, and exact retry appears only when safe. ArrowUp edits only your latest queued message before dispatch claims it.

Native resume when proved, bounded replay otherwise

Handoffs without the re-brief

Claude, Codex, and OpenCode resume by exact ID when proved. Copilot and Antigravity receive canonical user and peer context plus bounded own-history replay.

Typed references, restrained reactions

Quotes that keep agents aligned

Quote one message or several, react when a signal is enough, and keep collaboration readable without turning emoji into noise.

Review, approve, repair, roll back

Install tools from the conversation

Ask for a skill, MCP, hook, or role in chat. Review its targets, approve once, and keep repair or rollback on the same audited path.

Owned launch, drain, release, recovery

No mystery processes

Every CLI run carries its PID, exact start identity, host, and operation. VibeFlow only reclaims proved orphans, and live Windows Job Object validation runs in CI.

From one prompt to a proven change

Start naturally. Underneath, every handoff and action stays on one inspectable authority path.

  1. vf init

    Give every CLI the same brief

    Create shared repo context and install the vf skill where each supported CLI can use it.

  2. vf doctor --probe

    Know who is actually ready

    Probe installed engines, real round trips, and owned-process health before work is dispatched.

  3. vf orchestrate --yes

    Let the right agent take the turn

    Plan and dispatch work while a durable user-message queue keeps sends in order. Typed actions travel on their own audited authority path.

  4. vf verify

    Make done earn the word

    Hold completion until types, lint, tests, scope, evidence, and confidence all agree.

$ npx @magicpro97/vibeflow init
Seed shared context and sync the vf skill into every supported CLI
  ✓ claude
  ✓ codex
  ✓ copilot
  ✓ opencode
  ✓ antigravity

# inside Claude Code, Codex, Copilot, or OpenCode
> Refactor the auth middleware + vf
> /vf

Teach every CLI the same way of working

vf init gives each agent the shared workflow. Home keeps every later tool install visible and target-specific before anything changes.

One skill gives Claude Code, Codex, Copilot, OpenCode, and Antigravity the same VibeFlow workflow.

A capability request stays in the conversation, including exactly which CLIs will receive it.

Repair and rollback use the same visible path, so tool state cannot drift behind your back.

Learn more in the Skills system reference.

An optional fast judge, not another model

Turn on TypeSafe System One (Jev) and VibeFlow scores goal coverage in reviewer and verify flows. Turn it off and every gate behaves exactly as it did before: the integration is fail-open by construction.

Read the setup guide

No mystery processes

When a CLI finishes, hangs, or loses its owner, VibeFlow uses exact process identity before it releases or recovers anything.

Install locally bun install
Open the harness npx @magicpro97/vibeflow
Choose an ephemeral port vf ui --port 0
Probe engines vf doctor --probe
Repair owned CLI state vf doctor --fix
Verify before done vf verify

One fast toolchain

Bun 1.4 powers the CLI, local Home, tests, coverage, and landing while the shipped executable stays Node compatible.

A process has a fingerprint

Queue claims and CLI runs bind PID, start identity, host, operation, and nonce. Live or unprovable owners fail closed.

Drift gets a name

Doctor and status say manual, unsupported, orphaned, stale, or needs recovery instead of quietly pretending all is well.

One vocabulary, everywhere

Backend, CLI, and Home share frozen typed protocol authority and fail-closed guards instead of duplicate enums or raw wire literals.