ModelBound#

ModelBound is the source of truth for AI team skills, rules, system prompts, and knowledge bases. This skill connects OpenClaw to ModelBound's hosted MCP server so you can search and pull team context on demand instead of copy-pasting files.

Setup#

Required env:

• MODELBOUND_API_KEY — get one at https://modelbound.co/api-keys

Optional:

• MODELBOUND_TEAM_ID — scope all calls to a specific team
• MODELBOUND_ENDPOINT — defaults to https://mcp.modelbound.co

The MCP server is plain HTTPS JSON-RPC. No daemon, no shell-out.

Lean-first protocol#

ModelBound exposes 45+ tools. Do not advertise them all. Instead, call this small surface and let the server route:

ToolPurposemodelbound.listToolsList every available tool with its scope. Always call first if you don't know what's available.modelbound.callToolInvoke any tool by name. Use this for everything else.helpPlain-English description of a tool or workflow.search.allHybrid search across skills, rules, prompts, and corpora.search.summaryReranked top-N answer to a question.files.getFetch a specific file by ID or slug.skills.proposeDraftPropose an edit to a team-managed skill. Returns a review_url.gateway.setWorkspaceScope subsequent calls to one team / repo.

If a workflow below references a tool not on this list, route it through modelbound.callTool with the name from reference/tools.md.

Workflows#

1. Pull a team skill into the working directory#

1. gateway.setWorkspace with the user's team or repo slug.
2. search.all for the skill name (or skills.list via modelbound.callTool for an exhaustive list).
3. files.get to fetch the SKILL.md body.
4. Write it to ./.claude/skills/<name>/SKILL.md (or whatever location the user's IDE expects).

2. Answer a question from team knowledge#

1. search.summary with the user's question.
2. If the summary cites file IDs, follow up with files.get for the full text.
3. Cite the ModelBound URL the server returns — never fabricate one.

3. Read a specific file the user mentioned#

files.get with the slug or URL fragment. If ambiguous, search.all first to disambiguate.

4. Propose an edit back to ModelBound#

Never call skills.update directly. Always use skills.proposeDraft:

1. files.get the current version.
2. Compute the new content locally.
3. skills.proposeDraft with { skill_id, content, summary }.
4. Show the user the review_url it returns. A teammate approves the change in the ModelBound UI.

5. Discover more capabilities#

If the user asks for something this skill doesn't cover (evals, agents, corpora ingestion, MCP gateway), call modelbound.listTools and pick from the result. reference/tools.md has a categorized index.

Auth fallback#

If a call returns 401 unauthorized, tell the user:

Set MODELBOUND_API_KEY to a key from https://modelbound.co/api-keys, then re-run the request.

Don't try to recover by guessing or by hitting other endpoints.

Safety rules#

• HTTP only. Never shell out, never write to ModelBound files on disk other than under the user's IDE skill directory.
• Read tools ( search.*, files.get, *.list) are safe to call freely.
• Write tools ( skills.proposeDraft, corpus.upload, webhooks.create, etc.) require explicit user confirmation in chat first.
• The skill bans skills.update outright — propose drafts instead.

More#

• reference/tools.md — categorized index of all 45+ ModelBound MCP tools.
• reference/examples.md — three end-to-end JSON-RPC recipes.