MCP integration guide

Stream your context into AI dev tools

Reference PJT AI's project context directly from Claude Code · Cursor · Codex.

01What is MCP?#

MCP (Model Context Protocol) is a standard that lets AI models safely reference data in external systems. PJT AI provides a stdio MCP server (pjt-agent-mcp) in the @apexstack/agent package. Claude Code on your machine can query workspace data directly, and it is read-only.

Query data from Claude Code with these tools:

Tool	Description
`get_project`	Project basics (id or publicId `p_xxx`)
`list_project_tasks`	Task list (status filter available)
`get_task`	Task details
`get_document`	Document details (includes content)
`list_workspace_documents`	Workspace document list
`get_workspace_snapshot`	Workspace bundle (recent documents, etc.)

ⓘ

Separate from running work

MCP is for data queries only. It works independently of automatic task execution (pjt-agent start) and works without start. Authentication reuses your existing Claude Code agent token (pjta_…) as-is.

02Connect Claude Code#

Prerequisite: Node.js 18+ and Anthropic's claude CLI must be installed locally.

1. Issue a token

Issue a token (pjta_…) from workspace Settings → Claude Code agent → "Add agent". Right after issuing, the registration command is shown with the token already filled in (the token is shown only once). It uses the same token as the work-execution agent.

2. Install

bash

npm install -g @apexstack/agent

3. Register — one line (recommended)

Inject the token and URL via --env with no login, and it's a single line.

bash

claude mcp add pjt-workspace \
  --env PJT_AGENT_TOKEN=pjta_xxx \
  --env PJT_API_URL=https://api.pjt.ai \
  -- pjt-agent-mcp

Or register more briefly after login (saves a token file):

bash

pjt-agent login --token pjta_xxx
claude mcp add pjt-workspace -- pjt-agent-mcp

4. Verify it works

bash

$ claude
> Show me the recent documents in this workspace

[pjt-workspace] get_workspace_snapshot → 8 documents
[Claude] Here are the 8 most recent documents:
1. Spring Campaign Brief (06/07)
2. ...

✓

Connection successful

If pjt-workspace appears in claude mcp list and the tool-call logs above show up in the conversation, it's working.

03Cursor (coming soon)#

ⓘ

Launching in June

We'll publish the guide to coincide with Cursor IDE's official MCP support. For beta access, contact [email protected].

04Codex (coming soon)#

ⓘ

Launching in July

OpenAI Codex integration is scheduled to launch in the first week of July.

05Context scope#

By default, MCP exposes only the context of projects the user is a member of. Finer-grained control is available as follows:

Scope	Default	How to control
Per project	Membership-based	Workspace settings > Permissions
Sensitive fields	Includes budgets & quotes	Can be excluded in settings
Meeting notes	Full text exposed	Option to expose summary only

06Security · permissions#

Read-only — MCP cannot change or delete data (query only)
Reuses the Claude Code agent token (pjta_) — no separate key; revoke/reissue per agent
Token is shown only once at issuance — reissue if lost. Locally stored at ~/.pjt/agent.json (mode 0600)
Encrypted in transit and at rest (TLS 1.3 + AES-256); all access is recorded in audit logs
Not used for AI training — data ownership stays with the workspace
Enterprise can choose SSO/SAML and data residency (US · EU · KR)

MCP integration guide

Stream your context into AI dev tools

Reference PJT AI's project context directly from Claude Code · Cursor · Codex.

01What is MCP?#

Query data from Claude Code with these tools:

Tool	Description
`get_project`	Project basics (id or publicId `p_xxx`)
`list_project_tasks`	Task list (status filter available)
`get_task`	Task details
`get_document`	Document details (includes content)
`list_workspace_documents`	Workspace document list
`get_workspace_snapshot`	Workspace bundle (recent documents, etc.)

ⓘ

Separate from running work

02Connect Claude Code#

Prerequisite: Node.js 18+ and Anthropic's claude CLI must be installed locally.

1. Issue a token

2. Install

bash

npm install -g @apexstack/agent

3. Register — one line (recommended)

Inject the token and URL via --env with no login, and it's a single line.

bash

claude mcp add pjt-workspace \
  --env PJT_AGENT_TOKEN=pjta_xxx \
  --env PJT_API_URL=https://api.pjt.ai \
  -- pjt-agent-mcp

Or register more briefly after login (saves a token file):

bash

pjt-agent login --token pjta_xxx
claude mcp add pjt-workspace -- pjt-agent-mcp

4. Verify it works

bash

$ claude
> Show me the recent documents in this workspace

[pjt-workspace] get_workspace_snapshot → 8 documents
[Claude] Here are the 8 most recent documents:
1. Spring Campaign Brief (06/07)
2. ...

✓

Connection successful

If pjt-workspace appears in claude mcp list and the tool-call logs above show up in the conversation, it's working.

03Cursor (coming soon)#

ⓘ

Launching in June

We'll publish the guide to coincide with Cursor IDE's official MCP support. For beta access, contact [email protected].

04Codex (coming soon)#

ⓘ

Launching in July

OpenAI Codex integration is scheduled to launch in the first week of July.

05Context scope#

By default, MCP exposes only the context of projects the user is a member of. Finer-grained control is available as follows:

Scope	Default	How to control
Per project	Membership-based	Workspace settings > Permissions
Sensitive fields	Includes budgets & quotes	Can be excluded in settings
Meeting notes	Full text exposed	Option to expose summary only

06Security · permissions#

Read-only — MCP cannot change or delete data (query only)
Reuses the Claude Code agent token (pjta_) — no separate key; revoke/reissue per agent
Token is shown only once at issuance — reissue if lost. Locally stored at ~/.pjt/agent.json (mode 0600)
Encrypted in transit and at rest (TLS 1.3 + AES-256); all access is recorded in audit logs
Not used for AI training — data ownership stays with the workspace
Enterprise can choose SSO/SAML and data residency (US · EU · KR)