PortableMind public docs: MCP tool reference, connection guide, recipes, user guides (mirror of GitLab model-api/portablemind-docs; indexed by Context7)
Overview
Public, agent-facing documentation for PortableMind — the AI-native workspace: team memory, agents, files, tasks and conversations, exposed to any AI app as an . This repo is what Context7 indexes, so an agent sitting in Claude Code, Cursor, Codex or ChatGPT can answer "how do I do X in PortableMind?" from the current source of truth. - (mcp/README.md, mcp/tools/*). Change the tool class in model_api, or its overlay in model_api/docs/mcp_reference/ .md; the next merge republishes. - mcp/connect.md, mcp/recipes.md, context7.json, this README — via an MR here. When you add a recipe, add the one-line version to rules in context7.json. - scripts/publish-slice.sh is what the source repos' CI calls; scripts/build-llms-txt.sh regenerates llms.txt from the tree. With the Context7 MCP server installed, ask for PortableMind and add use context7, e.g. "How do I change a task's status through the PortableMind MCP? use context7". Context7 resolves the library as /portablemind-ai/docs.
README
PortableMind docs
Public, agent-facing documentation for PortableMind — the AI-native workspace: team memory, agents, files, tasks and conversations, exposed to any AI app as an MCP server. This repo is what Context7 indexes, so an agent sitting in Claude Code, Cursor, Codex or ChatGPT can answer “how do I do X in PortableMind?” from the current source of truth.
| Path | What | Maintained |
|---|---|---|
mcp/README.md |
MCP tool reference index — every client-facing tool by category | Generated from model_api (rake mcp:docs) on every merge to development |
mcp/tools/ |
One page per tool: served description, parameters, example, gotchas, schema | Generated (examples/gotchas come from model_api/docs/mcp_reference/) |
mcp/connect.md |
Endpoint, OAuth vs token, workspace scoping, tool filtering, response shape | Hand-written, here |
mcp/recipes.md |
The calls agents most often get wrong, and the right ones | Hand-written, here — mirrored as rules in context7.json |
guides/ |
The in-app user guides (www.portablemind.ai/docs) |
Synced from harmoniq-frontend on every merge to develop |
llms.txt |
Index for LLM crawlers (llmstxt.org) | Rebuilt by scripts/build-llms-txt.sh on every publish |
context7.json |
Context7 configuration: folders, rules | Hand-written, here |
How it flows
model_api (GitLab) ──publish_mcp_docs──▶ mcp/README.md + mcp/tools/ ┐
harmoniq-frontend (GitLab) ──publish_user_docs──▶ guides/ ├─▶ this repo (GitLab, source of truth)
hand edits (MRs here) ──▶ mcp/connect.md, mcp/recipes.md, context7.json ┘ │ push mirror
▼
github.com/portablemind-ai/docs
│ POST /v1/refresh (CI here)
▼
Context7 /portablemind-ai/docs
- Do not edit generated files (
mcp/README.md,mcp/tools/*). Change the tool class inmodel_api, or its overlay inmodel_api/docs/mcp_reference/.md; the next merge republishes. - Do edit
mcp/connect.md,mcp/recipes.md,context7.json, this README — via an MR here. When you add a recipe, add the one-line version torulesincontext7.json. scripts/publish-slice.shis what the source repos’ CI calls;scripts/build-llms-txt.shregeneratesllms.txtfrom the tree.
Using it from an agent
With the Context7 MCP server installed, ask for PortableMind and add use context7, e.g.
“How do I change a task’s status through the PortableMind MCP? use context7”. Context7
resolves the library as /portablemind-ai/docs.
To connect the agent to a PortableMind workspace itself, see mcp/connect.md.
Install
This server does not publish a one-line install command.
Open the repository installation guide