Skip to content

AgentWeave

The AgentWeave loop

Nothing appears out of thin air.

Run your AI factory in the open.

AgentWeave is a local app where AI coding agents work on your projects together, against specifications you approve. Every requirement, task, run, piece of evidence and merge is recorded and connected, so for anything in your repository you can say who made it, and why.

Get started How it fits together GitHub

pip install agentweave-ai
agentweave

The AgentWeave app: a project overview with what needs you, what is running, what merged, and each agent's state The AgentWeave app: a project overview with what needs you, what is running, what merged, and each agent's state

Why it exists

AgentWeave began as a way to get agents from different companies working together, each doing what it does best. Then the real problem showed. Work appeared out of thin air, and nobody could say who made it or why. A tool that lets agents loose on a codebase, and leaves you to work out afterwards what happened, does not give you control.

So AgentWeave became the place where you run the work in the open. Agents build from specifications you approve. Each task runs in its own git worktree. Evidence is recorded against the requirement it proves. Approval is a reviewed merge, and every step names who took it.

One thread through everything

In Goethe's Elective Affinities, every rope of the Royal Navy is spun with a red thread, so any piece of it can be traced back to where it was made. AgentWeave does the same for your software. Here is one thread from a real project, followed end to end:

REQUIREMENT FR-2 · rounding written by planner TASK Half-up rounding made by approval RUN builder · haiku own worktree EVIDENCE 3 tests passed commit 747084d REVIEW Approved by reviewer MERGE Into main by the Hub

Each of those is a record you can open: the requirement's history, the task's story, the run's transcript and trace, the evidence's command and output, the reviewer's verdict, and the commit that merged.

What you get

  • Specifications you approve

    Agents draft a change with requirements and scenarios. You check it and approve it, and approving creates the tasks; you build it when you choose.

    Specifications

  • Isolated, reviewed work

    Each task gets its own worktree and branch. A different agent reviews it, and approving it merges it into your main branch.

    Tasks and review

  • Traceability by default

    Task stories, requirement histories, evidence with its command and output, per-turn traces and a spend ledger.

    Traceability

  • Agents with a charter

    Claude Code, Codex and GitHub Copilot agents, each bound to a runner, a model and a charter that says what it answers for.

    Agents

  • Loops that finish

    Scheduled jobs, plus loops and flows with a purpose and a stop condition. Each one works a queue of tasks until the queue is done.

    Jobs, loops and flows

  • You stay in the loop

    Agents ask you questions and wait for the answer. Permission postures decide what an agent may do before it has to ask.

    Conversations

How it runs

AgentWeave is one local process, the Hub. It owns execution, state and identity, and it serves the app you work in. It runs on your machine, against directories on your disk, using the agent CLIs you have already signed in to. Nothing is hosted, and nothing is pushed anywhere unless you push it.

  • You work in the app: projects, agents, conversations, specs, tasks, jobs.
  • Agents are the CLIs the Hub launches. Each run gets a short-lived credential and a fixed set of tools for messages, tasks, questions, specs and evidence.
  • The CLI only starts, checks, stops and resets the local app. It also adds files to a project's knowledge vault.

Read How it fits together next, or go straight to the quick start.