Skip to main content
Glama

start_collaboration

Start a multi-agent collaboration where agents take turns responding in idea-building, debate, or conversation mode; returns a collaboration ID for managing the session.

Instructions

Start a collaboration where agents take turns responding; returns a collaboration ID. Default collaborate builds on ideas. Debate has agents[0] defend a position against challengers; waffle-house distills an idea through repeated attacks and rebuttals. Conversation is an informal chat from each agent's own situation, with short turns and no agenda or deliverable; topic is optional only in conversation mode. In conversation mode, put what you know about the agents' current situation in context (what was just built, what each agent last did). Private conversations default to grounding from workspace, memory and local activity; public conversations never get grounding, never open on the founder angle, and their directive does not ask agents to talk about the founder or whoever they report to. Pass prototype: true (not with conversation) to have agents show proposals as ephemeral prototypes that the other agents assess. Continue, nudge, inspect or end with the collaboration tools.

Tool call template

Replace each placeholder with a value matching the input schema. Only required parameters are included; add optional parameters as needed.

{
  "id": 1,
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "arguments": {
      "agents": "<agents>"
    },
    "name": "start_collaboration"
  }
}

Input Schema

NameRequiredDescriptionDefault
duelNoOptional duel id (duel-...) from ROUNDTABLE_HOME/duels to make the focus of this session: a scrubbed, size-capped brief of the duel (teams, artifacts, verdicts, tally, dissents) is added to the context every participant sees, as untrusted content to discuss. An unknown or invalid id is an error before any provider call.
modeNodebate: agents[0] is the proponent; every other agent is a challenger that stress-tests the proponent. waffle-house: agents[0] is the defender of an idea (the topic) and every other agent is an adversarial attacker (needs 2+ agents); each round every attacker attacks, then the defender answers the whole volley, and the session always ends on the defender's rebuttal. Speaking order is the same as collaborate; a session runs 1 + max_rounds x N agent turns. conversation: informal colleague chat grounded in each agent's own context, usually 2-6 sentences, no agenda or deliverable; topic may be omitted. Default: collaborate.
modelNoModel id. Defaults: claude-sonnet-5-5 (anthropic), gpt-5.6-luna (openai), zai-org/GLM-5.3 (together), qwen/qwen3-235b-a22b-instruct-2507 (replicate), qwen3:14b (ollama), default (openai_compatible). Model aliases: "opus" (claude-opus-5-5), "sonnet" (claude-sonnet-5-5), "haiku" (claude-haiku-5-5). Replicate accepts owner/name, owner/name:version, or a https://replicate.com/owner/name URL.
seatsNoOptional assignments keyed by exact agent identifier: {"designer":{"provider":"anthropic","model":"sonnet"}}. Each entry accepts agent, provider (a provider or explicit runtime name) and model.
topicNoRequired except in conversation mode, where it is an optional loose theme, not an agenda.
agentsYesArray of agent identifiers. Examples: ["cfo", "fpa"], ["product", "engineering", "design"]
publicNoDefault false. Only set true when the human explicitly asks for a public session. When true, the whole session is posted live to the configured public channel and is world-readable: the topic/agenda, every agent reply, every message you send with say/nudge, and the summary. The context argument is never posted, but agents may paraphrase it in their replies, so keep context generic. Never include personal information, credentials, or private code anywhere in a public session. The result includes public.url so the human can watch live.
contextNoOptional background for all agents. In conversation mode, share what you know about their current situation: what was just built, what each agent last did, and current problems.
base_urlNoOnly with provider "openai_compatible": base URL of an OpenAI-compatible /v1 endpoint (e.g. http://localhost:1234/v1). Overrides OPENAI_COMPATIBLE_BASE_URL for this session/call. OPENAI_COMPATIBLE_API_KEY, if set, is sent only to the OPENAI_COMPATIBLE_BASE_URL origin or one listed in ROUNDTABLE_OPENAI_COMPATIBLE_TRUSTED_ORIGINS; any other base URL gets no Authorization header and must be https (http only to loopback, or to private addresses with ROUNDTABLE_ALLOW_PRIVATE_BASE_URLS=1), never a cloud metadata or (without that setting) private address.
providerNoLLM provider. Omit or use auto for provider judgment (the router picks a healthy provider per seat and reports why in a routing block); otherwise the chosen provider is used as given. Options: auto; anthropic; openai; together (Together AI — any Together model id, e.g. zai-org/GLM-5.3 or Qwen/...); replicate (owner/name, owner/name:version, or https://replicate.com/owner/name); ollama (local, only when reachable); openai_compatible (any /v1/chat/completions server — set OPENAI_COMPATIBLE_BASE_URL or pass base_url).
groundingNoConversation only, default true: send workspace, agent-memory and local activity excerpts to the session's provider as background. Set false to disable. Always disabled in public conversations.
prototypeNoDefault false. Prototype mode: agents show proposals as ephemeral artifacts — a scaled-down page shown in a sandboxed frame on a localhost board, the code/IaC that would produce the proposal (never executed), a Mermaid diagram, or a text draft — and the other participants (in meetings: you) give verdicts; the proposer then revises, scraps or shortlists. Results include a prototypes block with the board URL: show it to the human. Prototypes are destroyed as soon as assessment completes. Do not relay the human's feedback on prototypes into the session until it ends; end_* lists outcomes and the shortlist for the human. Combinable with public (public posts get code, diagram links and screenshots of web prototypes, never the localhost board). Not supported with mode conversation.
connectorsNoOptional names of configured MCP connectors (roundtable config set connector) that every seat may call as tools during its turn. Only read-only tools are offered unless the connector's config lists more; tool results are untrusted data and are scrubbed. Seats on providers without tool support (Replicate) speak without them. An unknown connector is an error before any provider call.
max_roundsNoMaximum rounds of discussion (each round = each agent speaks once). Default: 5
page_checkNoCheck web prototypes in local headless Chrome when available (default true).
duel_sourceNoDefault true; only with duel. The duel brief is resent on every turn to every participant (a full brief can be about 20k tokens per turn). Set false to leave out artifact file contents and keep titles, file names and sizes, verdicts, tally and dissents.
debate_focusNoOptional: what the challenge should concentrate on (only used with mode debate).
auto_run_roundsNoIf set, automatically run this many rounds before returning. Otherwise returns after first agent speaks.
seat_connectorsNoOptional per-seat limits, used with connectors: a map from a seat (agent) name to the subset of connectors that seat may call, e.g. {"ada": ["figma"], "bob": ["jira"]}. A seat not listed gets all of the session's connectors; an empty list gives that seat none. Every connector named must also be in connectors, and every seat must be at the table; otherwise the call fails before anything starts.

Schema Changelog

Changes observed during successful MCP inspections.

  1. Changed19 schema fields changedv2.0.0
    • changedInput schema / properties / base_url / description
      Previous value: -"Only with provider \"openai_compatible\": base URL of an OpenAI-compatible /v1 endpoint (e.g. http://localhost:1234/v1). Overrides OPENAI_COMPATIBLE_BASE_URL for this session/call. OPENAI_COMPATIBLE_API_KEY, if set, is sent as the bearer token to whichever base URL is used."New value: +"Only with provider \"openai_compatible\": base URL of an OpenAI-compatible /v1 endpoint (e.g. http://localhost:1234/v1). Overrides OPENAI_COMPATIBLE_BASE_URL for this session/call. OPENAI_COMPATIBLE_API_KEY, if set, is sent only to the OPENAI_COMPATIBLE_BASE_URL origin or one listed in ROUNDTABLE_OPENAI_COMPATIBLE_TRUSTED_ORIGINS; any other base URL gets no Authorization header and must be https (http only to loopback, or to private addresses with ROUNDTABLE_ALLOW_PRIVATE_BASE_URLS=1), never a cloud metadata or (without that setting) private address."
    • addedInput schema / properties / connectors
      Added value: +{
      +  "description": "Optional names of configured MCP connectors (roundtable config set connector) that every seat may call as tools during its turn. Only read-only tools are offered unless the connector's config lists more; tool results are untrusted data and are scrubbed. Seats on providers without tool support (Replicate) speak without them. An unknown connector is an error before any provider call.",
      +  "items": {
      +    "maxLength": 32,
      +    "minLength": 1,
      +    "type": "string"
      +  },
      +  "maxItems": 8,
      +  "type": "array"
      +}
    • changedInput schema / properties / context / description
      Previous value: -"Optional background context or data to share with all agents."New value: +"Optional background for all agents. In conversation mode, share what you know about their current situation: what was just built, what each agent last did, and current problems."
    • addedInput schema / properties / debate_focus
      Added value: +{
      +  "description": "Optional: what the challenge should concentrate on (only used with mode debate).",
      +  "type": "string"
      +}
    • addedInput schema / properties / duel
      Added value: +{
      +  "description": "Optional duel id (duel-...) from ROUNDTABLE_HOME/duels to make the focus of this session: a scrubbed, size-capped brief of the duel (teams, artifacts, verdicts, tally, dissents) is added to the context every participant sees, as untrusted content to discuss. An unknown or invalid id is an error before any provider call.",
      +  "type": "string"
      +}
    • addedInput schema / properties / duel_source
      Added value: +{
      +  "description": "Default true; only with duel. The duel brief is resent on every turn to every participant (a full brief can be about 20k tokens per turn). Set false to leave out artifact file contents and keep titles, file names and sizes, verdicts, tally and dissents.",
      +  "type": "boolean"
      +}
    • addedInput schema / properties / grounding
      Added value: +{
      +  "description": "Conversation only, default true: send workspace, agent-memory and local activity excerpts to the session's provider as background. Set false to disable. Always disabled in public conversations.",
      +  "type": "boolean"
      +}
    • addedInput schema / properties / max_rounds / maximum
      Added value: +20
    • addedInput schema / properties / mode
      Added value: +{
      +  "description": "debate: agents[0] is the proponent; every other agent is a challenger that stress-tests the proponent. waffle-house: agents[0] is the defender of an idea (the topic) and every other agent is an adversarial attacker (needs 2+ agents); each round every attacker attacks, then the defender answers the whole volley, and the session always ends on the defender's rebuttal. Speaking order is the same as collaborate; a session runs 1 + max_rounds x N agent turns. conversation: informal colleague chat grounded in each agent's own context, usually 2-6 sentences, no agenda or deliverable; topic may be omitted. Default: collaborate.",
      +  "enum": [
      +    "collaborate",
      +    "debate",
      +    "waffle-house",
      +    "conversation"
      +  ],
      +  "type": "string"
      +}
    • changedInput schema / properties / model / description
      Previous value: -"Model id. Defaults: claude-opus-5 (anthropic), gpt-5.6-luna (openai), zai-org/GLM-5.3 (together), qwen/qwen3-235b-a22b-instruct-2507 (replicate), qwen3:14b (ollama), default (openai_compatible). Anthropic also accepts the aliases \"opus\" (claude-opus-5) and \"sonnet\" (claude-sonnet-5). Replicate accepts owner/name, owner/name:version, or a https://replicate.com/owner/name URL."New value: +"Model id. Defaults: claude-sonnet-5-5 (anthropic), gpt-5.6-luna (openai), zai-org/GLM-5.3 (together), qwen/qwen3-235b-a22b-instruct-2507 (replicate), qwen3:14b (ollama), default (openai_compatible). Model aliases: \"opus\" (claude-opus-5-5), \"sonnet\" (claude-sonnet-5-5), \"haiku\" (claude-haiku-5-5). Replicate accepts owner/name, owner/name:version, or a https://replicate.com/owner/name URL."
    • addedInput schema / properties / page_check
      Added value: +{
      +  "description": "Check web prototypes in local headless Chrome when available (default true).",
      +  "type": "boolean"
      +}
    • addedInput schema / properties / prototype
      Added value: +{
      +  "description": "Default false. Prototype mode: agents show proposals as ephemeral artifacts — a scaled-down page shown in a sandboxed frame on a localhost board, the code/IaC that would produce the proposal (never executed), a Mermaid diagram, or a text draft — and the other participants (in meetings: you) give verdicts; the proposer then revises, scraps or shortlists. Results include a prototypes block with the board URL: show it to the human. Prototypes are destroyed as soon as assessment completes. Do not relay the human's feedback on prototypes into the session until it ends; end_* lists outcomes and the shortlist for the human. Combinable with public (public posts get code, diagram links and screenshots of web prototypes, never the localhost board). Not supported with mode conversation.",
      +  "type": "boolean"
      +}
    • changedInput schema / properties / provider / description
      Previous value: -"LLM provider to use. Default: anthropic. Options: anthropic; openai; together (Together AI — any Together model id, e.g. zai-org/GLM-5.3 or Qwen/...); replicate (owner/name, owner/name:version, or https://replicate.com/owner/name); ollama (local, only when reachable); openai_compatible (any /v1/chat/completions server — set OPENAI_COMPATIBLE_BASE_URL or pass base_url)."New value: +"LLM provider. Omit or use auto for provider judgment (the router picks a healthy provider per seat and reports why in a routing block); otherwise the chosen provider is used as given. Options: auto; anthropic; openai; together (Together AI — any Together model id, e.g. zai-org/GLM-5.3 or Qwen/...); replicate (owner/name, owner/name:version, or https://replicate.com/owner/name); ollama (local, only when reachable); openai_compatible (any /v1/chat/completions server — set OPENAI_COMPATIBLE_BASE_URL or pass base_url)."
    • changedInput schema / properties / provider / enum
      Previous value: -[
      -  "anthropic",
      -  "openai",
      -  "together",
      -  "replicate",
      -  "ollama",
      -  "openai_compatible"
      -]New value: +[
      +  "auto",
      +  "anthropic",
      +  "openai",
      +  "together",
      +  "replicate",
      +  "ollama",
      +  "openai_compatible"
      +]
    • addedInput schema / properties / public
      Added value: +{
      +  "description": "Default false. Only set true when the human explicitly asks for a public session. When true, the whole session is posted live to the configured public channel and is world-readable: the topic/agenda, every agent reply, every message you send with say/nudge, and the summary. The context argument is never posted, but agents may paraphrase it in their replies, so keep context generic. Never include personal information, credentials, or private code anywhere in a public session. The result includes public.url so the human can watch live.",
      +  "type": "boolean"
      +}
    • addedInput schema / properties / seat_connectors
      Added value: +{
      +  "additionalProperties": true,
      +  "description": "Optional per-seat limits, used with connectors: a map from a seat (agent) name to the subset of connectors that seat may call, e.g. {\"ada\": [\"figma\"], \"bob\": [\"jira\"]}. A seat not listed gets all of the session's connectors; an empty list gives that seat none. Every connector named must also be in connectors, and every seat must be at the table; otherwise the call fails before anything starts.",
      +  "patternProperties": {
      +    "^.+$": {
      +      "items": {
      +        "maxLength": 32,
      +        "minLength": 1,
      +        "type": "string"
      +      },
      +      "maxItems": 8,
      +      "type": "array"
      +    }
      +  },
      +  "type": "object"
      +}
    • addedInput schema / properties / seats
      Added value: +{
      +  "additionalProperties": true,
      +  "description": "Optional assignments keyed by exact agent identifier: {\"designer\":{\"provider\":\"anthropic\",\"model\":\"sonnet\"}}. Each entry accepts agent, provider (a provider or explicit runtime name) and model.",
      +  "patternProperties": {
      +    "^.+$": {
      +      "additionalProperties": false,
      +      "properties": {
      +        "agent": {
      +          "type": "string"
      +        },
      +        "model": {
      +          "minLength": 1,
      +          "type": "string"
      +        },
      +        "provider": {
      +          "description": "Provider or explicit configured runtime name.",
      +          "pattern": "^[a-z][a-z0-9_-]{0,31}$",
      +          "type": "string"
      +        }
      +      },
      +      "type": "object"
      +    }
      +  },
      +  "type": "object"
      +}
    • changedInput schema / properties / topic / description
      Previous value: -"The topic or question for agents to discuss."New value: +"Required except in conversation mode, where it is an optional loose theme, not an agenda."
    • changedInput schema / required
      Previous value: -[
      -  "agents",
      -  "topic"
      -]New value: +[
      +  "agents"
      +]
  2. First observedv0.1.0

TDQS

A3.9/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Annotations only declare readOnly=false, destructive=false, openWorld=true, idempotent=false; the description goes well beyond by disclosing that public sessions are posted live and world-readable, that public conversations never get grounding and never open on the founder angle, that prototypes are destroyed once assessment completes and must not be relayed mid-session, and that conversation must carry the agents' current situation in context. These are genuine operational constraints an agent could not derive from the annotations. It leaves the auto_run_rounds/return-timing behavior (returns after the first agent speaks unless set) to the schema rather than restating it.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness3/5

Is the description appropriately sized, front-loaded, and free of redundancy?

Front-loaded with purpose and the collaboration ID, then mode semantics, then constraints, which is a sensible order. But it is a dense run-on paragraph that partly restates the mode enum descriptions already in the schema, and the grounding/public/prototype clauses are packed into two long sentences that are harder to scan than a bulleted list would be.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

This is a 19-parameter tool with nested objects and no output schema, and the description covers the decisions that matter most: which mode, what topic requires, grounding and public exposure, and prototype mode. Return behavior is only partly set — it mentions the collaboration ID and points to end_* for outcomes, without describing the initial result payload — but given the schema coverage the remaining gaps are modest.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Schema description coverage is 100%, so baseline is 3; the description still adds cross-parameter meaning the schema states only locally — topic being optional only in conversation mode, prototype not being combinable with conversation, and the instruction to include situational knowledge in context for conversation mode. It does not address max_rounds, provider/model, connectors or seat_connectors, which the schema already covers adequately.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose4/5

Does the description clearly state what the tool does and how it differs from similar tools?

States a specific verb and resource ("Start a collaboration where agents take turns responding") and names the return value (collaboration ID), which distinguishes it cleanly from continue_collaboration, end_collaboration and get_collaboration_transcript. It does not differentiate itself from other starters in the sibling set (start_meeting, start_duel, start_local_meeting), so the agent must infer that boundary.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

Gives explicit selection criteria across the four modes — collaborate builds on ideas, debate has agents[0] defend against challengers, waffle-house distills through repeated attacks, conversation is an informal chat with no deliverable — which is real when-to-use guidance. It also routes follow-up actions to the collaboration tools. It gives no exclusion telling the agent when to prefer start_meeting or start_duel instead.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.