# Client workspaces for consultants, with scoped agent access

> Keep each client in its own private Cozy workspace and connect your AI agents through a hosted MCP with read, write and folder-scoped credentials.

Updated 2026-11-09. Canonical: https://cozy.cloud/use-cases/consultants

> Create one private Cozy workspace per client so engagements stay apart. For your own day-to-day work, connect your agent once to `https://cozy.cloud/mcp` and name the client's workspace on each call. When an agent must never see another client's notes, connect it to that client's workspace URL instead, with OAuth or a bearer credential scoped to read, wiki-write or source-write and an optional folder prefix.

## The problem with one pile of memory

Consultants and agencies run several engagements at once. Meeting notes, decisions, deliverables and contracts pile up per client. When you add an AI agent, the easy path is to point it at one big pile of notes. That works until the agent quotes Client A's pricing in Client B's draft.

The fix is a boundary that the tool enforces, not one you remember to respect. In Cozy, that boundary is the workspace. Every account gets "My wiki", and you can create more private workspaces to keep projects apart. One workspace per client gives each engagement its own pages, sources and agent notes.

## How the setup works

1. Create a workspace for the client and name it after the engagement.
2. Keep it to yourself. Cozy is single-player, so clients and colleagues are not invited in. Send deliverables to the client through the channels you already use.
3. Connect your agent to `https://cozy.cloud/mcp` and approve it on Cozy's consent page. One connection reaches all your workspaces. The agent calls `cozy_workspaces` to list them and passes the client's workspace ID or exact name as the `workspace` argument. Without it, tools use My wiki.
4. For an agent that should only ever reach one client, use that workspace's own URL, `https://cozy.cloud/wiki-mcp/WORKSPACE_UUID`, shown under Agents in the workspace. OAuth on that URL grants only that workspace. For scripts and headless agents, create a bearer credential there with the narrowest scope that does the work.
5. Put any bearer credential in the agent's config as an environment variable.

Cozy runs no model itself. You bring your own agent, such as Claude, Codex or Cursor, and the workspace is the memory it reads and writes. See [MCP server](/mcp-server) for the transport details.

## Scoped credentials

An agent connected with OAuth gets no more access than you have in each workspace, capped by the scopes it was granted (wiki:read, wiki:write, sources:write). Connected agents are listed under Connectors → Agents, where you can disconnect them.

Bearer credentials for a single workspace are created in Agents. They are random, shown once, stored as SHA-256 digests, revocable, and they expire within 90 days. Each one has a scope.

| Scope | What the agent can do |
| --- | --- |
| read | Read and search pages and sources |
| wiki-write | Write and edit curated pages |
| source-write | Add evidence under sources/ |

You can also add a folder-prefix grant, so a credential only touches paths under a given prefix. A credential cannot give an agent more than you may do yourself.

Agents cannot see connector credentials or permanently purge content. That matters for client work, because an agent can never remove an audit trail.

### Example credentials per engagement

- A research agent gets read plus source-write, so it can file evidence but not rewrite curated pages.
- A drafting agent gets read plus wiki-write, limited to a `deliverables/` prefix.
- A reporting agent gets read only.

## Connecting an agent

For your own interactive work, add the account URL once. In Claude Code:

```bash
claude mcp add --transport http --scope user cozy https://cozy.cloud/mcp
```

Then run `/mcp` to sign in. In Codex, run `codex mcp add cozy --url https://cozy.cloud/mcp`, which opens a browser to sign in. Ask the agent to name the client's workspace on every call, and stay in one workspace per task.

For a headless agent that should only reach one client, use the workspace URL and a bearer credential. For Codex, add this to the config:

```toml
[mcp_servers.cozy]
url = "https://cozy.cloud/wiki-mcp/WORKSPACE_UUID"
bearer_token_env_var = "COZY_MCP_TOKEN"
```

For Claude Code, use a `.mcp.json` file:

```json
{"mcpServers":{"cozy":{"type":"http","url":"https://cozy.cloud/wiki-mcp/WORKSPACE_UUID","headers":{"Authorization":"Bearer ${COZY_MCP_TOKEN}"}}}}
```

Keep one project directory per client, each with its own `.mcp.json` and its own token. Opening the client's directory then connects the right workspace by default. OAuth also works on a workspace URL, configured with no Authorization header, and grants only that workspace. Setup guides: [Claude Code](/connect/claude-code) and [Codex](/connect/codex).

## Keeping client knowledge organized

Each workspace has four folders: `wiki/` for curated pages with an `index.md`, `sources/` for evidence from connectors and importers, `skills/` for editable SKILL.md files, and `agents/` for agent notes and handoffs. Content under `sources/` is treated as untrusted data, which is useful when a client email contains instructions you do not want an agent to follow.

Connectors include Granola, Italic, Gmail, Google Calendar, Google Drive, Outlook mail, Outlook Calendar, Notion, Slack for selected channels, Linear, Ramp and Custom MCP. You can also push files and exports through the ingest API. A stable document id means a re-push updates the same source.

Pages link to each other with wikilinks, and Cozy builds a graph from Markdown conventions on every save. Page types such as client, stakeholder or decision come from frontmatter, so an agent can ask for a client page and its typed neighbors with `cozy_context`. The [typed links](/typed-links) page covers the syntax.

## Quality and handover

Edits require an expected revision, so two agents editing the same page produce a conflict instead of a silent overwrite. Deletion is a tombstone. The last 50 revisions of a page can be read and restored, and history is audited.

The health check reports broken links, stale pages whose cited source changed, conflicts, orphans and unsourced pages, with a score and a prioritized list of fixes. Run it before you hand a deliverable to a client. See [wiki health](/wiki-health).

At the end of an engagement you can revoke every credential for that workspace and disconnect its agents without touching your other clients.

## Where other tools are stronger

Notion is a better fit if the client already lives in Notion and wants databases, views and comments, or if you and the client need to edit pages together. Cozy is single-player. Notion's hosted MCP uses OAuth, and per [Notion's developer docs](https://developers.notion.com/docs/mcp) it works on Notion's block and database model. Notion supports bearer-token headless access only on Business and Enterprise. Cozy stores plain Markdown, JSON and CSV, which is easier to move out, but it has no database views. See [Cozy vs Notion](/compare/notion).

Cozy search is lexical. There are no embeddings or vector database, so agents find pages by words, links and graph structure, not by semantic similarity.

## Limits to know

- Curated files can be up to 90,000 bytes. Imported sources can be up to 1,500,000 bytes.
- Browser upload allows 1 MB per file and 25 files per selection.
- An agent connected to the account URL can reach all your workspaces. Use a workspace URL when an agent must be limited to one client.
- Cozy is operated by Coral Sourcing LLC and is not affiliated with Cozy Cloud (cozy.io) or Cozi. Privacy details are on the [privacy page](/privacy).

Check your own client contracts before putting confidential material into any hosted service.

## Frequently asked questions

### Can I give each client their own workspace?

Yes. Create a private workspace per client. Only you and the agents you authorize can access it. Your agent reaches them all through https://cozy.cloud/mcp and picks one with the workspace argument. Each workspace also has its own MCP URL under Agents when you want agent access kept to one client.

### How do I stop an agent from seeing other clients' notes?

Connect it to that client's workspace URL instead of the account URL, with OAuth or a bearer credential created in that workspace. OAuth on a workspace URL grants only that workspace. Credentials can be limited to read, wiki-write or source-write scopes and to a folder prefix, and they never exceed your own access.

### How long do agent credentials last?

Bearer credentials expire within 90 days and can be revoked at any time in Agents. OAuth connections renew with refresh tokens and can be disconnected at any time under Connectors → Agents.

### Can an agent delete client data permanently?

No. Agents cannot see connector credentials or permanently purge content. Deletions are tombstones, and the last 50 revisions can be restored.

### Which agents work with a client workspace?

Any MCP client that supports Streamable HTTP, including [Claude Code](/connect/claude-code), [Codex](/connect/codex) and [Cursor](/connect/cursor). Cozy runs no model itself, so you bring your own agent.
