Route big coding tasks to another agent, and review the diff before it lands.

pactify sends large coding tasks from Claude Code or your terminal to codex, kimi, opencode or agy, runs each one as a background job, in its own git worktree when you ask for one, and hands you the result to review, apply or discard.

curl -fsSL https://pactify.dev/install.sh | sh
Read the docs

macOS and Linux. Then install and log in to the harness CLI you want to use.

pactify in 60 seconds, without sound: Claude hands a real task to codex, reviews the diff and reruns the tests, then applies the change.
A real job: Claude Code asked for one change, pactify routed it to codex, and the result came back for review.

Claude Code

/delegate Add pagination to the tasks API

  1. codex Default Succeeded in 2m 50s
  2. kimi Supported Moonshot AI
  3. opencode Supported Your configured provider
  4. agy Experimental Google 
✓ codex succeededHarness: codex (minimal config) · CLI: 0.162.1Model: gpt-6.1-sol → gpt-6.1-solAccess: writeDuration: 2m50sTokens: in 236749 · out 6245Quota: 7-day window: 5% used, resets 10-30 18:54Tool calls: 21Changes: 5 files · +311/-11Worktree: ~/.pactify/jobs/20261024-003121-af0e95/wt- README.md (+19/-1)- handlers.go (+38/-1)- handlers_test.go (+176/-2)- store.go (+16/-7)- store_test.go (+62/-0)

How a delegation runs

  1. Write the brief

    Ask Claude Code to delegate, or pipe a brief into pactify sub run. The brief states the goal, the scope and the acceptance commands, because the other agent starts without your conversation.

  2. It runs in the background

    Each job gets an id, a log and, when you ask for one, its own git worktree; the Claude Code plugin always asks. Keep working, check status, wait, cancel or resume. When a harness hits its quota, pactify can fall back to the next one you configured.

  3. Review what came back

    The result names the model that actually ran, the tokens and quota it used, and every changed file. Read the full diff and rerun your checks in the worktree.

  4. Apply, rework or discard

    Apply copies the change into your checkout without committing and stops on conflicts. Resume sends rework to the same session and worktree. Discard removes it.

Worth delegating, and when not to

Each delegation costs the other agent 20 to 50 thousand tokens before it starts on your task. Send work that touches several files, needs exploration and test runs, or would take you more than ten minutes. Keep small edits in your own session.

What you can check

  • The requested model and the model that actually ran, side by side.
  • Changes isolated in a worktree based on your current commit.
  • A read-only mode on every harness, built on that harness's own controls.
  • No telemetry and no pactify server. Your harness CLI talks to its own provider.
How each harness is isolated

Install

CLI

curl -fsSL https://pactify.dev/install.sh | sh

Claude Code plugin

/plugin marketplace add https://pactify.dev/marketplace.json/plugin install pactify@pactify

Also available through go install, and as plugins for Codex, Copilot and Cursor. All install options

Running several coding agents across a team?

We are exploring a Team edition for visibility and policy across vendors. Everything pactify does on your own machine stays free.

Tell us what your team needs