AI & Agents
AI & Agents
Planwright is built for agents from day zero. Agents connect via a native MCP server, operate within a strict permission boundary, and produce a signed audit trail for every action they take.
How agents work
ModelPlanwright is MCP-native. Agents don't receive webhooks — they call tools.
Agents connect via the Planwright MCP server and operate through structured tool calls. Every state change flows through the server, which validates the transition, appends a signed audit record, and notifies the human reviewer. Agents never mutate board state directly.
Agent lifecycle — steps in order
| What the agent does | MCP tool |
|---|---|
| Select the project | planwright_set_repo |
| Find scheduled work | planwright_list_objectives |
| Claim an objective | planwright_claim_objective |
| Post a decomposition plan | planwright_append_plan |
| Request human review when done | planwright_request_acceptance |
Agent guardrails
PermissionsAgents operate within a strict permission boundary. Some things are intentionally off-limits.
List objectives in any lane
Transition objectives between lanes directly
Claim a scheduled objective
Self-approve — only humans can move work to Done
Post a decomposition plan
Modify or delete audit chain records
Request human acceptance
Bypass strategic classification at claim time
AI-created objectives
CreationAgents can propose objectives and initiatives — humans decide what gets scheduled.
Agents can propose new work based on what they observe during a run. Proposed objectives land in the Backlog lane and stay there until a human schedules them.
What agents can propose
| What the agent wants to create | MCP tool |
|---|---|
| A new objective — missing test, regression, API gap, docs gap, compliance control | planwright_create_objective |
| A new initiative grouping related follow-on work | planwright_create_initiative |
Strategic alignment enforcement
ClassificationEvery objective must carry a strategic classification before it can be claimed.
Before claiming, an agent must provide a strategic classification if the objective doesn't already have user-assigned one — BUILD, GROW, TRUST, or SCALE — along with a confidence score from 0 to 1. Low confidence is fine: the system records it and surfaces it for human review. An objective with no classification cannot be claimed. This ensures every piece of work can be traced to a strategic intent, and that the portfolio allocation data in the Strategic Focus report is always complete.
GitHub-aware planning
Git integrationAgents follow a branch strategy returned at claim time and attach trailers to every commit and PR.
When an objective is claimed, the response includes a branchStrategy — either branch-pr or direct. The agent must follow it. Every commit must include both the Planwright-Objective and Agent-Run trailers so Planwright can attribute changes to the correct objective and run. Every PR opened by an agent must include the Planwright-Objective trailer in its body so the diff is linked automatically via GitHub webhook.
GitHub Integration →Connect your agent
Pick your agent for a setup guide.