Website · Install · Docs · Pipelines · Releases

One agent session runs out of context on a big task, then grades its own work. Convoy carries the task across many fresh sessions, checkpointed and resumable, then has independent models score the branch and fix the gaps until it reaches 90/100. On OpenCode, Claude Code, Codex, or any mix.

Why Convoy

• Goes the distance. A big task runs as phases. Each one starts with a fresh context and ends in a commit, so a run can detach, resume, and keep going for hours. Running Convoy →
• Proves the result. Two independent scorers grade the branch against a fixed rubric, and fix rounds target only the gaps they found until it reaches 90/100. Quality and goals →
• Bring your harness. Every step runs on OpenCode, Claude Code, or Codex (experimental), mixed per step, with advisors from any of them. A model is spelled the same everywhere; the step's runner picks the harness. Runners →
• Same rules for every agent. Per-step policies (read-only, verify, write) apply to every tool call. Agents can't commit or push. Safety →
• Everything stays local. Logs, transcripts, reports, and costs live next to your Git history. Worktrees and OpenSpec changes keep each piece of work isolated. Worktrees → · OpenSpec →

Get started

Requires macOS or Linux, Git, and at least one installed, authenticated runner: OpenCode, Claude Code, or Codex. Bun is bundled in the release binary.

Warning

Codex support is experimental. It runs on Codex's app-server protocol, which OpenAI itself labels experimental, and it has had less real-world use than OpenCode and Claude Code. Expect rough edges and breaking changes between releases, and check the branch before you merge it. Please report issues. Codex in Convoy →

curl -fsSL https://github.com/Inakitajes/convoy/releases/latest/download/install.sh | sh

Run convoy in your project to open the terminal UI, or start from the CLI:

convoy --prompt-file prd.md       # implement, measure, and fix with full-cycle
convoy -p review                  # report-only review of the current branch
convoy -p claude-full-cycle       # the full cycle on Claude Code alone

Convoy also runs headless, so a script or another agent can supervise it. convoy skill install teaches Claude Code, OpenCode, and Codex sessions to compose a pipeline for a task, launch it detached, and follow it with convoy runs <id> --wait:

convoy skill install              # then ask your agent: "run a Convoy review of this branch", or type /convoy-review

See installation and updates, provider setup, driving Convoy from a script or agent, and agent sessions.

Pick a pipeline

Pipeline	Purpose
full-cycle · default	Implement, audit, polish, and test, then score and fix gaps until the branch reaches 90/100.
implement	Implement and audit, then write a run recap. No scoring loop.
ship	Sync with the base, review and fix findings, then reach the quality target.
review	Produce findings and a verified score without changing code.
fixer	Prove supplied findings with regression tests, fix them, and verify the result.
hunter	Audit six specialty tracks across five models, then reconcile the findings.
claude-full-cycle	The full cycle on Claude Code alone, scored by Opus and Sonnet.
claude-review	A report-only Claude Code review, synthesized by Opus.

Every pipeline is YAML you can copy and change, or write in a file of its own and run with --pipeline-file. convoy pipelines lists what can run. Models, steps, and stopping rules →

Documentation

Guide	Covers
Getting started	Requirements, installation, and upgrades
Pipelines	Built-ins, models, and migration from earlier variants
Runners	OpenCode, Claude Code, and Codex: models, routes, auth, billing, and parity
Quality and goals	Scoring rubric and improve/measure loops
Worktrees	Home, isolation, spin, and close
OpenSpec	Spec-driven runs and the artifact reader
Running Convoy	CLI, TUI, headless runs, history, and resume
Agent sessions	The Convoy skill for Claude Code, OpenCode, and Codex
Configuration	Agents, advisors, prompts, hooks, and attachments
Permissions and safety	Approval policies, commits, compaction, and publication
Models and providers	Authentication and model gateways
Development	Building, testing, and the source layout

License

MIT © Iñaki Tajes