Studio tools

The multi-agent studio: roster, board, feedback, decisions, loops, messages and playtests.

26 tools in the studio category. Badges: read-only tools never change the project; mutating tools are undoable and attributed; open-world tools reach outside the engine and MCP clients ask before running them.

Tool	Summary
studio_overview	Start here when working with the studio (a team of AI agents making this game).
studio_agent_define	Create a studio agent or update one (by id or name; only the fields you pass change).
studio_agent_list	The studio roster: each agent's id (use it for assignees and @mentions), name, role, discipline, focus, live status, open tasks and usage.
studio_agent_brief	Everything needed to be a roster member: its system prompt (role mission, focus, persona, standing instructions, memory, the team, studio etiquette) and the engine tools it may use with their access (allow | ask).
studio_agent_remove	Remove an agent from the studio (deletes agents/<id>.agent.json; its open tasks become unassigned).
studio_team_template	Spawn a whole team of specialists from a template (agents that already exist by id are updated): starter_crew (5 generalists), indie_trio (3), aaa_strike_team (22: director, producer, five designers, four programmers, five artists, audio, writing, QA, two playtester personas, critic), narrative_team, qa_squad (QA lead, three playtester personas, critic, fixer), art_team, audio_team.
studio_memory	Your long-term memory as a studio agent (kept in your profile and shown at the start of every task): note a fact worth remembering (conventions, decisions, the human's preferences, unfinished work), forget an outdated note by number, or list notes.
studio_usage_report	Record tokens an agent spent with its own model, for studio budgets and cost tracking (runners that call models outside the engine — the editor crew, custom scripts — report each request here).
studio_presence	Show what a roster member is doing right now in the editor's Studio panel (live status and activity next to its avatar): working with a one-line activity when you start something, idle when you finish.
studio_task_create	Put a task on the studio board.
studio_task_update	Move a task across the board (backlog → todo → doing → review → done, or dropped), reassign it, edit it, or add a comment.
studio_task_list	The studio board.
studio_task_claim	Claim a task and start it (assignee = you, status = doing).
studio_feedback_submit	File feedback about the game as a playtester, critic, agent or human: category (fun, difficulty, clarity, visuals, audio, performance, bug, narrative, accessibility), severity, a one-line summary, details, what it's about (target), and evidence {playtest: "P-3", captures: [...], metrics: {...}, positions: [[x,y,z]], repro: [steps]}.
studio_feedback_list	Feedback with each item's director verdict (act / drop / defer / merge_into), rationale and measured effect (metrics before/after).
studio_feedback_update	Change a feedback item: status (e.g.
studio_decide	The director's (or producer's) verdict on a feedback item, visible to everyone with its rationale: act (creates tasks — give each a title, acceptance criteria and an assignee — or links existing task_ids), drop (won't do, say why), defer (later), or merge_into a duplicate.
studio_loop_define	Define a studio loop: a repeatable cycle of stages with stop conditions, stored in studio/loops/<name>.loop.json.
studio_loop_start	Start a run of a loop.
studio_loop_advance	Report the results of a loop stage's assignments and move the loop on: reports [{agent, report, signoff?, usage?}].
studio_loop_status	A loop's state: status, iteration, current stage, pending assignments (with prompts), per-iteration history (stage reports, playtests, filed feedback, verified effects) and metric trends.
studio_loop_stop	Stop a running loop (pending assignments are dropped).
studio_message_send	Talk to the team so the back-and-forth is visible: post in a channel (#general by default; any name creates a channel), address agents with to or @mentions in the text (@id, @role, @discipline, @all), reply in a thread with reply_to, and reference a task/feedback.
studio_inbox	Messages for you (addressed to you, mentioning you, or broadcast in #general), oldest first; unread only by default, and reading marks them read.
playtest_run	Play the game with a bot in a sandbox copy of the scene (the scene you edit is untouched) and get a report: completion rate, time-to-goal, deaths and their causes, fails, damage, stuck periods, objectives, coverage, frame cost, screenshots at notable moments and a top-down heatmap (returned as an image).
playtest_compare	Before/after comparison of two playtests (ids like P-3): every metric with its delta and whether it got better or worse (direction-aware: fewer deaths is better, higher completion is better).

studio_overview

Studio overview read-only

Start here when working with the studio (a team of AI agents making this game). Returns the roster with live status, the task board counts, feedback waiting for a decision, recent director decisions with measured effects, loops and their state, the latest playtest metrics, and token/cost usage. include_catalog=true also lists roles, disciplines, team templates and loop templates.

Argument	Type	Required	Description	Values
include_catalog	boolean		Also return roles, team templates and loop templates

Tool callCLIMCP request

studio_overview {"include_catalog": true}

skywalker call studio_overview '{"include_catalog": true}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_overview",
    "arguments": {
      "include_catalog": true
    }
  }
}

studio_agent_define

Define agent mutating

Create a studio agent or update one (by id or name; only the fields you pass change). An agent is a specialist: a role (e.g. level_designer, gameplay_programmer, lighting_artist, playtester, critic — see studio_overview include_catalog), a focus (what exactly it owns: "enemy AI", "onboarding"), a persona, standing instructions, a model, an autonomy level (observe = read-only, ask = changes need human approval, autonomous), per-tool-category permissions (allow|ask|off), and who it reports to. Saved as agents/.agent.json (shareable, git-friendly). Playtesters take persona knobs in playtest {policy, reaction_time, skill, curiosity, patience}. Example: {name:"Mira", role:"level_designer", focus:"secret areas", focus_tags:["secrets"], reports_to:"nimbus"}.

Argument	Type	Required	Description	Values
id	string		Agent id to update (default: derived from name)
name	string		Display name
role	string		Role	creative_director art_director producer systems_designer level_designer narrative_designer economy_designer ux_designer gameplay_programmer ai_programmer graphics_programmer tools_programmer environment_artist lighting_artist character_artist vfx_artist technical_artist ui_artist sound_designer composer writer qa_lead playtester critic
discipline	string		Discipline (default: from the role)	direction production design engineering art audio writing qa
focus	string		What this agent specializes in
focus_tags	string[]		Short focus tags
persona	string		Personality / voice
mission	string		Replaces the role's built-in mission
instructions	string		Standing instructions: style guides, constraints, conventions
provider	string		anthropic | openai | mock | an editor provider name
model	string		Model id (empty = provider default, e.g. claude-opus-5-5)
autonomy	string		How freely it may change the game	observe ask autonomous
permissions	object		Per tool category:
reports_to	string		Agent id of its lead
color	string		Avatar color #rrggbb
face	string		Avatar face	happy focused dreamy wink determined curious
max_rounds	integer		Max tool rounds per task (default 40)
memory	string[]		Replace its long-term memory notes
playtest	object		Playtester persona:

Tool callCLIMCP request

studio_agent_define {"id": "level_designer", "name": "Mira", "role": "level_designer", "focus": "layout and pacing", "reports_to": "creative_director"}

skywalker call studio_agent_define '{"id": "level_designer", "name": "Mira", "role": "level_designer", "focus": "layout and pacing", "reports_to": "creative_director"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_agent_define",
    "arguments": {
      "id": "level_designer",
      "name": "Mira",
      "role": "level_designer",
      "focus": "layout and pacing",
      "reports_to": "creative_director"
    }
  }
}

studio_agent_list

List agents read-only

The studio roster: each agent's id (use it for assignees and @mentions), name, role, discipline, focus, live status, open tasks and usage. Filter by discipline or role; include_profiles returns full profiles (persona, instructions, permissions, memory).

Argument	Type	Required	Description	Values
discipline	string		Only this discipline
role	string		Only this role
include_profiles	boolean		Full profiles

Tool callCLIMCP request

studio_agent_list {"include_profiles": true}

skywalker call studio_agent_list '{"include_profiles": true}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_agent_list",
    "arguments": {
      "include_profiles": true
    }
  }
}

studio_agent_brief

Agent brief read-only

Everything needed to be a roster member: its system prompt (role mission, focus, persona, standing instructions, memory, the team, studio etiquette) and the engine tools it may use with their access (allow | ask). Use it to run a studio agent in any harness: e.g. give a Claude Code / Codex subagent this system prompt, let it call only the listed tools, and connect it as "/" (or pass as: "") so its work is attributed. loop_member=true withholds loop and roster control tools.

Argument	Type	Required	Description	Values
agent	string	yes	Agent id or name
loop_member	boolean		Running inside a loop stage

Tool callCLIMCP request

studio_agent_brief {"agent": "{{STUDIO_ID}}", "loop_member": true}

skywalker call studio_agent_brief '{"agent": "{{STUDIO_ID}}", "loop_member": true}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_agent_brief",
    "arguments": {
      "agent": "{{STUDIO_ID}}",
      "loop_member": true
    }
  }
}

studio_agent_remove

Remove agent mutating destructive

Remove an agent from the studio (deletes agents/.agent.json; its open tasks become unassigned).

Argument	Type	Required	Description	Values
agent	string	yes	Agent id or name

Tool callCLIMCP request

studio_agent_remove {"agent": "aurora"}

skywalker call studio_agent_remove '{"agent": "aurora"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_agent_remove",
    "arguments": {
      "agent": "aurora"
    }
  }
}

studio_team_template

Spawn a team mutating

Spawn a whole team of specialists from a template (agents that already exist by id are updated): starter_crew (5 generalists), indie_trio (3), aaa_strike_team (22: director, producer, five designers, four programmers, five artists, audio, writing, QA, two playtester personas, critic), narrative_team, qa_squad (QA lead, three playtester personas, critic, fixer), art_team, audio_team. Override provider/model/autonomy for everyone with defaults.

Argument	Type	Required	Description	Values
template	string	yes	Team template	starter_crew indie_trio aaa_strike_team narrative_team qa_squad art_team audio_team
defaults	object		Fields applied to every member, e.g.

Tool callCLIMCP request

studio_team_template {"template": "aaa_strike_team"}

skywalker call studio_team_template '{"template": "aaa_strike_team"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_team_template",
    "arguments": {
      "template": "aaa_strike_team"
    }
  }
}

studio_memory

Agent memory mutating

Your long-term memory as a studio agent (kept in your profile and shown at the start of every task): note a fact worth remembering (conventions, decisions, the human's preferences, unfinished work), forget an outdated note by number, or list notes.

Argument	Type	Required	Description	Values
action	string	yes	What to do	note forget list
text	string		The note (for note)
number	integer		Note number, 1-based (for forget)
agent	string		Whose memory (default: you)
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_memory {"action": "note", "agent": "nimbus", "text": "The art director prefers warm rim light on hero props."}

skywalker call studio_memory '{"action": "note", "agent": "nimbus", "text": "The art director prefers warm rim light on hero props."}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_memory",
    "arguments": {
      "action": "note",
      "agent": "nimbus",
      "text": "The art director prefers warm rim light on hero props."
    }
  }
}

studio_usage_report

Report usage mutating

Record tokens an agent spent with its own model, for studio budgets and cost tracking (runners that call models outside the engine — the editor crew, custom scripts — report each request here). Tool calls are counted automatically.

Argument	Type	Required	Description	Values
agent	string		Agent id (default: you)
model	string		Model id (prices the tokens)
input_tokens	integer		Uncached input tokens
output_tokens	integer		Output tokens
cache_read_tokens	integer		Cached input tokens read
cache_write_tokens	integer		Input tokens written to the cache
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_usage_report {"agent": "level_designer", "model": "claude-sonnet-5-5", "input_tokens": 12000, "output_tokens": 3400, "cache_read_tokens": 80000}

skywalker call studio_usage_report '{"agent": "level_designer", "model": "claude-sonnet-5-5", "input_tokens": 12000, "output_tokens": 3400, "cache_read_tokens": 80000}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_usage_report",
    "arguments": {
      "agent": "level_designer",
      "model": "claude-sonnet-5-5",
      "input_tokens": 12000,
      "output_tokens": 3400,
      "cache_read_tokens": 80000
    }
  }
}

studio_presence

Set presence mutating

Show what a roster member is doing right now in the editor's Studio panel (live status and activity next to its avatar): working with a one-line activity when you start something, idle when you finish. External harnesses (the Python agent layer, scripts, Claude Code subagents) call it so their agents appear live like the editor's crew. Example: {status: "working", activity: "T-4 widen the lava bridge", as: "mira"}.

Argument	Type	Required	Description	Values
status	string	yes	Status	working idle waiting blocked
activity	string		One line: what you are doing (empty when idle)
agent	string		Whose presence (default: you)
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_presence {"status": "working", "activity": "T-4 widen the bridge", "as": "level_designer"}

skywalker call studio_presence '{"status": "working", "activity": "T-4 widen the bridge", "as": "level_designer"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_presence",
    "arguments": {
      "status": "working",
      "activity": "T-4 widen the bridge",
      "as": "level_designer"
    }
  }
}

studio_task_create

Create task mutating

Put a task on the studio board. Good tasks are small and checkable: a clear title, acceptance criteria (how we know it's done), a discipline, and an assignee (agent id, or "@role" to pick the least busy agent with that role). Link the feedback it addresses so the effect can be measured. Example: {title:"Widen the lava bridge", acceptance:["bridge 3 m wide","deaths ≤ 1 per run"], assignee:"@level_designer", feedback:["F-2"]}.

Argument	Type	Required	Description	Values
title	string	yes	Short imperative title
description	string		Context and details
acceptance	string[]		Acceptance criteria, one per item
discipline	string		Discipline	direction production design engineering art audio writing qa
assignee	string		Agent id or @role
priority	string		Priority (default normal)	low normal high critical
status	string		Initial status (default todo)	backlog todo doing review done dropped
depends_on	string[]		Task ids that must be done first
feedback	string[]		Feedback ids this task addresses
links	object		Other links:
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_task_create {"title": "Widen the second canyon gap's landing", "discipline": "design", "priority": "high", "acceptance": ["Bots clear the gap in 3 of 3 runs"]}

skywalker call studio_task_create '{"title": "Widen the second canyon gap'\''s landing", "discipline": "design", "priority": "high", "acceptance": ["Bots clear the gap in 3 of 3 runs"]}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_task_create",
    "arguments": {
      "title": "Widen the second canyon gap's landing",
      "discipline": "design",
      "priority": "high",
      "acceptance": [
        "Bots clear the gap in 3 of 3 runs"
      ]
    }
  }
}

studio_task_update

Update task mutating

Move a task across the board (backlog → todo → doing → review → done, or dropped), reassign it, edit it, or add a comment. Marking the last task of a feedback item done moves that feedback to "fixed" (the next verification playtest then measures the effect). Moving to doing/done checks dependencies.

Argument	Type	Required	Description	Values
task	string	yes	Task id, e.g. T-4
status	string		New status	backlog todo doing review done dropped
assignee	string		Agent id (empty string unassigns)
priority	string		Priority	low normal high critical
title	string		Title
description	string		Description
acceptance	string[]		Acceptance criteria (replaces the list)
discipline	string		Discipline
depends_on	string[]		Replace dependencies
links	object		Links to add
comment	string		Comment for the thread (what changed, what's left)
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_task_update {"task": "T-12", "status": "review", "comment": "Landing widened by 1.5 m; playtest attached."}

skywalker call studio_task_update '{"task": "T-12", "status": "review", "comment": "Landing widened by 1.5 m; playtest attached."}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_task_update",
    "arguments": {
      "task": "T-12",
      "status": "review",
      "comment": "Landing widened by 1.5 m; playtest attached."
    }
  }
}

studio_task_list

Task board read-only

The studio board. Filter by status (or "open" = backlog/todo/doing/review), assignee, discipline, or mine=true for your own tasks. Returns tasks with acceptance criteria, links and comment counts; the text is a compact kanban.

Argument	Type	Required	Description	Values
status	string		A status, or "open" / "all" (default all)
assignee	string		Agent id
discipline	string		Discipline
mine	boolean		Only tasks assigned to you
limit	integer		Max tasks (default 100)
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_task_list {"status": "open", "assignee": "level_designer"}

skywalker call studio_task_list '{"status": "open", "assignee": "level_designer"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_task_list",
    "arguments": {
      "status": "open",
      "assignee": "level_designer"
    }
  }
}

studio_task_claim

Claim task mutating

Claim a task and start it (assignee = you, status = doing). Without task, claims your highest-priority todo task, or else an unassigned todo task in your discipline. External agents identify with as or a client name ending in /.

Argument	Type	Required	Description	Values
task	string		Task id (default: pick the next one for you)
agent	string		Claim for this agent (default: you)
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_task_claim {"as": "level_designer"}

skywalker call studio_task_claim '{"as": "level_designer"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_task_claim",
    "arguments": {
      "as": "level_designer"
    }
  }
}

studio_feedback_submit

Submit feedback mutating

File feedback about the game as a playtester, critic, agent or human: category (fun, difficulty, clarity, visuals, audio, performance, bug, narrative, accessibility), severity, a one-line summary, details, what it's about (target), and evidence {playtest: "P-3", captures: [...], metrics: {...}, positions: [[x,y,z]], repro: [steps]}. The director triages it with studio_decide. Pass a fingerprint to deduplicate recurring findings (a repeat bumps occurrences instead of filing a duplicate).

Argument	Type	Required	Description	Values
category	string	yes	Category	fun difficulty clarity visuals audio performance bug narrative accessibility
severity	string		Severity (default medium)	low medium high critical
summary	string	yes	One line: what's wrong or great
details	string		What happened, where, why it matters
target	string		Entity, area, system or asset it is about
evidence	object		{playtest, captures, metrics, positions, repro, trace}
fingerprint	string		Dedup key for recurring issues
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_feedback_submit {"category": "difficulty", "severity": "high", "summary": "Players fall at the second canyon gap", "target": "Canyon Gap 2", "as": "playtester"}

skywalker call studio_feedback_submit '{"category": "difficulty", "severity": "high", "summary": "Players fall at the second canyon gap", "target": "Canyon Gap 2", "as": "playtester"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_feedback_submit",
    "arguments": {
      "category": "difficulty",
      "severity": "high",
      "summary": "Players fall at the second canyon gap",
      "target": "Canyon Gap 2",
      "as": "playtester"
    }
  }
}

studio_feedback_list

Feedback read-only

Feedback with each item's director verdict (act / drop / defer / merge_into), rationale and measured effect (metrics before/after). status: open, needs_decision (open + regressed), active (accepted, in_progress, fixed), a single status, or all (default).

Argument	Type	Required	Description	Values
status	string		Filter (default all)
category	string		Category	fun difficulty clarity visuals audio performance bug narrative accessibility
limit	integer		Max items, newest first (default 50)

Tool callCLIMCP request

studio_feedback_list {"status": "open", "category": "bug"}

skywalker call studio_feedback_list '{"status": "open", "category": "bug"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_feedback_list",
    "arguments": {
      "status": "open",
      "category": "bug"
    }
  }
}

studio_feedback_update

Update feedback mutating

Change a feedback item: status (e.g. verified after a critic's re-review, regressed to reopen its tasks), severity, category, summary, target, more evidence, or a comment for its history. Decisions themselves go through studio_decide.

Argument	Type	Required	Description	Values
feedback	string	yes	Feedback id, e.g. F-3
status	string		New status	open accepted dropped deferred in_progress fixed verified regressed
severity	string		Severity	low medium high critical
category	string		Category	fun difficulty clarity visuals audio performance bug narrative accessibility
summary	string		Summary
details	string		Details
target	string		Target
evidence	object		Evidence to merge in
comment	string		Note for the history
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_feedback_update {"feedback": "F-7", "status": "verified", "comment": "Deaths at the gap went from 4 to 0."}

skywalker call studio_feedback_update '{"feedback": "F-7", "status": "verified", "comment": "Deaths at the gap went from 4 to 0."}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_feedback_update",
    "arguments": {
      "feedback": "F-7",
      "status": "verified",
      "comment": "Deaths at the gap went from 4 to 0."
    }
  }
}

studio_decide

Decide on feedback mutating

The director's (or producer's) verdict on a feedback item, visible to everyone with its rationale: act (creates tasks — give each a title, acceptance criteria and an assignee — or links existing task_ids), drop (won't do, say why), defer (later), or merge_into a duplicate. The current playtest metrics are stored as the baseline, so the next verification playtest measures the effect (optionally against explicit targets like {"deaths":{"max":1}}); a regression reopens the tasks. Only direction/production agents (or humans) may decide. Example: {feedback:"F-2", verdict:"act", rationale:"Unfair spike", tasks:[{title:"Telegraph the lava", assignee:"@lighting_artist", acceptance:["glow visible from 10 m"]}]}.

Argument	Type	Required	Description	Values
feedback	string	yes	Feedback id
verdict	string	yes	Verdict	act drop defer merge_into
rationale	string	yes	Why — shown next to the feedback
tasks	object[]		Tasks to create (act)
task_ids	string[]		Existing tasks that address it (act)
merge_into	string		Feedback id it duplicates (merge_into)
targets	object		What success means: {metric: {min, max}}
metrics_before	object		Baseline metrics (default: latest playtest)
redecide	boolean		Override an earlier decision
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_decide {"feedback": "F-7", "verdict": "act", "rationale": "Unfair difficulty spike early in the level.", "targets": {"deaths": {"max": 1}}}

skywalker call studio_decide '{"feedback": "F-7", "verdict": "act", "rationale": "Unfair difficulty spike early in the level.", "targets": {"deaths": {"max": 1}}}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_decide",
    "arguments": {
      "feedback": "F-7",
      "verdict": "act",
      "rationale": "Unfair difficulty spike early in the level.",
      "targets": {
        "deaths": {
          "max": 1
        }
      }
    }
  }
}

studio_loop_define

Define loop mutating

Define a studio loop: a repeatable cycle of stages with stop conditions, stored in studio/loops/.loop.json. Start from a template (playtest_fix_verify, art_pass_with_critic, balance_tuning, vertical_slice_sprint, bug_bash) and override goal/stop, or write stages yourself. A stage is either kind "playtest" (the engine plays the game with bots: playtest {policy, runs, seconds}, file_feedback, verify_fixed) or kind "agents" (assignees: ["@role", "@discipline", "agent-id", "@a|@b" fallback], instruction template with {{goal}} {{inputs}} {{open_feedback}} {{my_tasks}} {{metrics}} {{roster}} {{iteration}}, parallel, only_with_tasks, gate {skip_if: no_open_feedback | no_todo_tasks | not_first_iteration | ..., approval: "human"}). stop: {max_iterations, metric_targets {completion_rate:{min:0.8}}, director_signoff, token_budget, time_budget_minutes}. delete=true removes it.

Argument	Type	Required	Description	Values
name	string		Loop name (default: the template's)
template	string		Start from this template	playtest_fix_verify art_pass_with_critic balance_tuning vertical_slice_sprint bug_bash
goal	string		What the loop is trying to achieve
description	string		Description
stages	object[]		Stages (replace the template's)
stop	object		Stop conditions (merged over the template's)
delete	boolean		Remove the loop

Tool callCLIMCP request

studio_loop_define {"template": "playtest_fix_verify", "goal": "Players finish the canyon run without unfair deaths"}

skywalker call studio_loop_define '{"template": "playtest_fix_verify", "goal": "Players finish the canyon run without unfair deaths"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_loop_define",
    "arguments": {
      "template": "playtest_fix_verify",
      "goal": "Players finish the canyon run without unfair deaths"
    }
  }
}

studio_loop_start

Start loop mutating

Start a run of a loop. The engine runs playtest stages itself and returns the first agent stage's assignments: [{agent, stage, prompt, tasks}]. Execute each assignment as that agent (in-editor crew, the headless runner skywalker studio run, or your own subagents with actor "/"; parallel=true means they may run at the same time), then report back with studio_loop_advance. Override goal or max_iterations for this run.

Argument	Type	Required	Description	Values
loop	string	yes	Loop name
goal	string		Goal for this run
max_iterations	integer		Iteration limit for this run

Tool callCLIMCP request

studio_loop_start {"loop": "playtest_fix_verify", "goal": "Players finish the canyon run without unfair deaths", "max_iterations": 3}

skywalker call studio_loop_start '{"loop": "playtest_fix_verify", "goal": "Players finish the canyon run without unfair deaths", "max_iterations": 3}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_loop_start",
    "arguments": {
      "loop": "playtest_fix_verify",
      "goal": "Players finish the canyon run without unfair deaths",
      "max_iterations": 3
    }
  }
}

studio_loop_advance

Advance loop mutating

Report the results of a loop stage's assignments and move the loop on: reports [{agent, report, signoff?, usage?}]. When every pending assignment has reported (or complete_stage=true), the loop runs the following stages — playtests, verification, effect measurement — until it needs agents again, waits for approval, or ends on its stop conditions. Returns the new status with the next assignments. Directors sign off with signoff=true (or a report starting with SIGNOFF). approve=true/false answers a human approval gate.

Argument	Type	Required	Description	Values
loop	string	yes	Loop name
reports	object[]		Reports from the stage's agents
approve	boolean		Answer an approval gate
signoff	boolean		Director/producer sign-off for this iteration
complete_stage	boolean		Finish the stage even if some agents did not report
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_loop_advance {"loop": "playtest_fix_verify", "reports": [{"agent": "nimbus", "report": "…"}]}

skywalker call studio_loop_advance '{"loop": "playtest_fix_verify", "reports": [{"agent": "nimbus", "report": "…"}]}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_loop_advance",
    "arguments": {
      "loop": "playtest_fix_verify",
      "reports": [
        {
          "agent": "nimbus",
          "report": "…"
        }
      ]
    }
  }
}

studio_loop_status

Loop status read-only

A loop's state: status, iteration, current stage, pending assignments (with prompts), per-iteration history (stage reports, playtests, filed feedback, verified effects) and metric trends. Without a loop name, lists every loop and the available templates.

Argument	Type	Required	Description	Values
loop	string		Loop name (omit to list)
history	boolean		Include iteration history (default true)

Tool callCLIMCP request

studio_loop_status {"loop": "playtest_fix_verify", "history": true}

skywalker call studio_loop_status '{"loop": "playtest_fix_verify", "history": true}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_loop_status",
    "arguments": {
      "loop": "playtest_fix_verify",
      "history": true
    }
  }
}

studio_loop_stop

Stop loop mutating

Stop a running loop (pending assignments are dropped).

Argument	Type	Required	Description	Values
loop	string	yes	Loop name
reason	string		Why

Tool callCLIMCP request

studio_loop_stop {"loop": "playtest_fix_verify"}

skywalker call studio_loop_stop '{"loop": "playtest_fix_verify"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_loop_stop",
    "arguments": {
      "loop": "playtest_fix_verify"
    }
  }
}

studio_message_send

Send message mutating

Talk to the team so the back-and-forth is visible: post in a channel (#general by default; any name creates a channel), address agents with to or @mentions in the text (@id, @role, @discipline, @all), reply in a thread with reply_to, and reference a task/feedback. Example: {text:"@lighting_artist the lava needs a stronger glow, see F-2", feedback:"F-2"}.

Argument	Type	Required	Description	Values
text	string	yes	Message (supports @mentions)
to	string[]		Recipients: agent ids or @roles
channel	string		Channel (default general)
reply_to	string		Message id to reply to (threads)
task	string		Related task id
feedback	string		Related feedback id
kind	string		Structured message kind for agent harnesses (default chat): request, inform, handoff, result, approval_request, approval_answer, ...
data	object		Structured payload for agent harnesses (the text stays the human-readable summary)
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_message_send {"text": "@nimbus the canyon fix is ready for review", "channel": "level", "task": "T-12"}

skywalker call studio_message_send '{"text": "@nimbus the canyon fix is ready for review", "channel": "level", "task": "T-12"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_message_send",
    "arguments": {
      "text": "@nimbus the canyon fix is ready for review",
      "channel": "level",
      "task": "T-12"
    }
  }
}

studio_inbox

Inbox read-only

Messages for you (addressed to you, mentioning you, or broadcast in #general), oldest first; unread only by default, and reading marks them read. Check it when you start a task. channel reads a whole channel; agent reads someone else's inbox without marking.

Argument	Type	Required	Description	Values
agent	string		Whose inbox (default: you)
unread_only	boolean		Only unread (default true)
channel	string		Read this channel instead ("*" = every channel)
thread	string		Read a whole thread instead (root message id); never marks read
after	string		With channel or thread: only messages newer than this id
limit	integer		Max messages (default 30)
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

studio_inbox {"unread_only": true}

skywalker call studio_inbox '{"unread_only": true}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "studio_inbox",
    "arguments": {
      "unread_only": true
    }
  }
}

playtest_run

Playtest mutating

Play the game with a bot in a sandbox copy of the scene (the scene you edit is untouched) and get a report: completion rate, time-to-goal, deaths and their causes, fails, damage, stuck periods, objectives, coverage, frame cost, screenshots at notable moments and a top-down heatmap (returned as an image). Policies: goal_seeker (walks to entities tagged "goal"/"objective", avoids "hazard"/"enemy"), explorer (covers unexplored space), random, scripted (script: [{t: seconds, hold|release|press|click|event: ...}] — the same arguments as sim_input). persona {reaction_time, skill, curiosity, patience} shapes the bot; playtester agents use their own persona by default. runs>1 aggregates several seeds. Games should tag the player "player" and emit "death" / "fail" / "damage" / "objective" events. Saved under studio/playtests//; findings are returned for filing as feedback.

Argument	Type	Required	Description	Values
policy	string		Bot policy (default goal_seeker)	goal_seeker explorer random scripted
seconds	number		Simulated seconds per run (default 30, max 600)
runs	integer		Runs with consecutive seeds (default 1, max 20)
seed	integer		First seed (default 1)
persona	object		{reaction_time: s, skill: 0..1, curiosity: 0..1, patience: s}
controls	object		Key mapping {up, down, left, right, jump} (default w/s/a/d/space)
script	object[]		Input timeline (scripted policy)
player	integer | string		Player entity (default: tagged "player")
goal_radius	number		Meters from a goal that count as reaching it (default 1)
screenshots	boolean		Capture screenshots (default true)
screenshot_interval	number		Also capture every N seconds
max_screenshots	integer		Screenshot cap (default 8)
label	string		Label for the report
agent	string		Use this playtester's persona
include_heatmap	boolean		Return the heatmap image (default true)
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

playtest_run {"policy": "goal_seeker", "runs": 3, "seconds": 45, "label": "before"}

skywalker call playtest_run '{"policy": "goal_seeker", "runs": 3, "seconds": 45, "label": "before"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "playtest_run",
    "arguments": {
      "policy": "goal_seeker",
      "runs": 3,
      "seconds": 45,
      "label": "before"
    }
  }
}

playtest_compare

Compare playtests mutating

Before/after comparison of two playtests (ids like P-3): every metric with its delta and whether it got better or worse (direction-aware: fewer deaths is better, higher completion is better). Pass feedback to record the result as that feedback's measured effect: improved → verified, regressed → regressed (its tasks reopen). Timing metrics (*_ms, est_fps) depend on the machine's load, so they only decide the verdict for performance feedback or when a target names them; they are always listed.

Argument	Type	Required	Description	Values
before	string	yes	Earlier playtest id
after	string	yes	Later playtest id
feedback	string		Record the effect for this feedback item
as	string		Act as this roster member (agent id). For external agents whose MCP client name isn't "<client>/<agent id>" (e.g. Claude Code subagents playing a role). Ignored when you already are one.

Tool callCLIMCP request

playtest_compare {"before": "P-3", "after": "P-4", "feedback": "F-2"}

skywalker call playtest_compare '{"before": "P-3", "after": "P-4", "feedback": "F-2"}' --project my_game

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "playtest_compare",
    "arguments": {
      "before": "P-3",
      "after": "P-4",
      "feedback": "F-2"
    }
  }
}