Set the outcome.
The team runs the play.
Specialized agents propose every step: spec, architecture, plan, QA. Your team approves each decision before it becomes the next.
The AI proposes. Your team decides.
Illustrative - orchestration is Planned
The problem
A software project isn't one task. It's a chain of decisions that compound.
TaskForce turns each into a checkpoint: proposed, reviewed, approved before it becomes context for the next.
Inside a run
One outcome. Every safeguard, in the open.
A run reads your context, proposes each step, waits for a human decision, hands the approved work to your coding agent, then measures itself against the forecast.
Grounded
It starts from what you already decided
Before proposing anything, the run reads your decisions, constraints and conventions.
Go deeper on MemoryHandoff
Approved plans go straight to your coding agent
TaskForce hands the full approved context to the agent you already use. It builds on a new branch and opens a pull request.
How the local runner worksHuman-in-the-loop
Every checkpoint is a decision, not a setting
The agent proposes. Your team approves it or sends it back - nothing moves on its own.
Predict & recalibrate
Coming nextIt predicts, then checks itself
Each step forecasts effort, risk and cost, then compares the forecast with what actually shipped.
Illustrative - orchestration is Planned
Local runner
BetaAssign the issue. Your agent opens the pull request.
Pick Claude Code as the assignee and a small runner on your computer takes the issue. It starts your own Claude Code, signed in with your own account, in an isolated branch. Nothing is merged until a person reviews it.
- 01
Assign
Choose Claude Code in the issue's assignee menu. The run waits in TaskForce.
- 02
Claim
The runner on your computer fetches it. TaskForce never connects to your machine.
- 03
Build
Claude Code works in a fresh worktree and reads the issue through the TaskForce MCP server.
- 04
Review
The runner pushes the branch and opens the pull request. The issue moves to “In review by AI”.
What the agent can touch
| Your repository | An isolated worktree on a new branch. Your checkout and your current branch are never touched. |
|---|---|
| Your machine | A closed list of tools: read, edit and commit. No free shell, no network, no push. |
| TaskForce | The rights of the person who delegated, narrowed to reading the workspace and writing that project's issues. Never a deletion. |
| GitHub | Nothing. The runner pushes and opens the pull request once the agent is done. |
The text of an issue is handed to the agent as data, never as instructions, and the scope above caps what a planted instruction could reach. No repository linked? The agent works in a disposable folder and answers with a comment on the issue: no git, no pull request.
Set it up yourself: in TaskForce, open Settings, then Agents, and create your runner. You get its credentials and the commands to start it. It runs on your computer with Node, git, the GitHub CLI and Claude Code, on your own subscription or API key. TaskForce never sees your Claude credentials.
Set up your runnerThe team
One outcome. Three specialists. Seven checkpoints.
Responsibility is split the way a strong software org does: each agent proposes in its lane, and your team approves before it moves.
- -Frame the problem - users, outcome, DoD
- -Draft the product spec - stories, acceptance
- -Propose the approach - architecture, risks
- -Write the API contract - endpoints, payloads
- -Break down & sequence the work
- -QA and sign-off against acceptance
Run it your way
Your models. Your numbers. One human gate.
Run every agent on the model you choose (local at zero cost or hosted for depth) and keep the whole run legible: seven checkpoints, three specialists, one decision-maker.
Your models, your call
Per runLocal through Ollama at zero model cost, or a hosted model when you want more depth.
Why orchestration matters
The model is replaceable.
The orchestration isn't.
Models and agents will change. The hard part is coordinating the chain of decisions: preserving context, exposing trade-offs, keeping humans accountable.
Grounded
Built from your architecture, decisions and conventions.
Specialized
Each agent stays within a defined responsibility.
Governed
Every transition is a human decision, not a setting.
Your models
Local via Ollama at zero model cost, or hosted.
Describe the outcome. Let the team work the path.
Your team decides at every gate. Your coding agent builds it.
Orchestration is Planned. The board, approvals and memory ship today.