Put an agent on the job. Keep a person on the hook.
Connect Claude, ChatGPT, Cursor, VS Code or your own code to os.construction through MCP or the osc command. It reads what you can read, drafts what you could draft, and stops at every approval.
@os-construction/cli 0.2.0 needs Node.js 24 or newer and has no runtime dependencies. The command is osc; the MCP connector is osc mcp. Searching the catalog needs no account.
Pick your OS
npx -y @os-construction/cli operations search --query "change events" --limit 5# or install it oncenpm i -g @os-construction/cliosc version
npx -y @os-construction/cli operations search --query "change events" --limit 5# or install it once (use npm.cmd / osc.cmd if script policy blocks .ps1 shims)npm i -g @os-construction/cliosc version
Connect
Pick your assistant. Copy. Done.
Every snippet matches Account > AI assistants in the app, which also shows which path your workspace can use right now.
Claude Desktop today with a token; claude.ai web, desktop and mobile through the hosted connector.
Works today · read-only preview
Hosted connector (claude.ai, Desktop, mobile)
In Claude open Settings > Connectors and choose Add custom connector. On Team and Enterprise plans an Owner adds it under Organization settings > Connectors, then each member chooses Connect.
Name it os.construction and paste https://app.os.construction/mcp. Keep the authentication options Claude detects.
Choose Connect, sign in to os.construction and choose Allow access.
Works today
Claude Desktop with a personal token
In os.construction open Account > AI assistants > Personal API tokens. Name it (say Claude on my laptop), keep Read-only (recommended), pick 7, 30 or 90 days and choose Create token. It is shown once.
In Claude Desktop choose Settings > Developer > Edit config and add the server below to claude_desktop_config.json. Node.js 24 or newer must be installed.
Restart Claude Desktop, then ask: "Search os.construction for RFI operations and list the open RFIs on project 123."
The config file lives at ~/Library/Application Support/Claude/claude_desktop_config.json on macOS and %APPDATA%\Claude\claude_desktop_config.json on Windows.
ChatGPT connects only to hosted MCP servers, so it waits for the hosted connection.
Works today · read-only preview
ChatGPT app (developer mode)
In ChatGPT open Settings > Apps. If your plan requires it, turn on developer mode under the advanced settings.
Create an app, name it os.construction, paste https://app.os.construction/mcp and choose OAuth authentication.
Sign in to os.construction when prompted and choose Allow access.
One command registers the local connector; the hosted path signs in through your browser.
Works today · read-only preview
Hosted connection
Add the server over HTTP.
Run /mcp in Claude Code, choose os-construction and authenticate in your browser.
Hosted connectionShell
claude mcp add --transport http os-construction https://app.os.construction/mcp
Works today
Claude Code with a personal token
In os.construction open Account > AI assistants > Personal API tokens. Name it (say Claude on my laptop), keep Read-only (recommended), pick 7, 30 or 90 days and choose Create token. It is shown once.
Register the local connector. Prefer your shell or secret manager over pasting the token into history (see the OS tabs).
Start Claude Code and run /mcp to check that os-construction is connected.
Same mcpServers block as Claude Desktop today; a static client ID for the hosted path.
Works today · read-only preview
Hosted connection
Add this to ~/.cursor/mcp.json or .cursor/mcp.json. Cursor needs the CLIENT_ID line because it does not publish a client identity os.construction can verify.
In Cursor's MCP settings choose Connect and allow access in your browser.
In os.construction open Account > AI assistants > Personal API tokens. Name it (say Claude on my laptop), keep Read-only (recommended), pick 7, 30 or 90 days and choose Create token. It is shown once.
Add the block below to ~/.cursor/mcp.json (or .cursor/mcp.json in a project you do not commit).
Open Cursor's MCP settings and check that os-construction shows its tools.
In os.construction open Account > AI assistants > Personal API tokens. Name it (say Claude on my laptop), keep Read-only (recommended), pick 7, 30 or 90 days and choose Create token. It is shown once.
Add this to .vscode/mcp.json. VS Code asks for the token when the server starts and keeps it out of the file.
Start the server from the file or the MCP view, then use Copilot in agent mode.
Replace 10 with your company ID and 123 with a project you can open. Both are in the app URL: app.os.construction/companies/10/projects/123/....
Launch the stdio connector from any MCP client, or drive osc as a subprocess and parse its JSON.
Works today
Bring your own agent
Any MCP client (Claude Agent SDK, OpenAI Agents SDK, LangChain MCP adapters, your own): launch npx -y @os-construction/cli mcp as a stdio server with OS_CONSTRUCTION_API_TOKEN in its environment. You get os_search_operations, os_describe_operation and os_read (plus os_write when you set OS_CONSTRUCTION_MCP_ALLOW_WRITES=true on a read-and-write token).
No MCP? Run osc as a subprocess. Every command writes exactly one JSON envelope to stdout and a stable exit code, so branching is trivial.
Unattended jobs should use a service account key instead of a person's token. See the service account recipe.
Pick your OS
import { execFile } from 'node:child_process';import { promisify } from 'node:util';const run = promisify(execFile);// One JSON envelope on stdout; exit code says what happened (0 ok, 3 auth, 4 forbidden, 9 unknown write outcome...)async function osc(...args) { try { const { stdout } = await run('npx', ['-y', '@os-construction/cli', ...args], { env: process.env }); return { exit: 0, ...JSON.parse(stdout) }; } catch (e) { if (!e.stdout) throw e; return { exit: e.code, ...JSON.parse(e.stdout) }; }}const found = await osc('operations', 'search', '--query', 'rfis', '--method', 'GET', '--limit', '5');console.log(found.data.operations.map(o => `${o.id} ${o.summary}`));
import json, subprocessdef osc(*args): p = subprocess.run(["npx", "-y", "@os-construction/cli", *args], capture_output=True, text=True) return p.returncode, json.loads(p.stdout) # exactly one envelope per callcode, env = osc("operations", "search", "--query", "change events", "--limit", "5")for op in env["data"]["operations"]: print(op["id"], op["method"], op["path"])
What can it do
2,341 operations. Generated, not guessed.
Straight from the operation catalog in the os.construction source (revision 3eff1ae8265e): 1,064 reads and 1,277 writes across 60 modules. 222 API definitions are deliberately left out (non-JSON bodies, unsafe paths, missing contracts). The npm package 0.2.0 carries revision 2b7478125d87 with 2,191 operations; newer ones arrive with each CLI release. osc coverage shows what yours has.
RFIs, submittals, drawings, specs, meetings, documents and the rest of the project record.
GETRead current budget facts for the request's cost codes/rest/v1.0/projects/{project_id}/purchase_requests/{id}/budget_contextRead
POSTPOST /rest/v1.0/projects/{project_id}/purchase_requests/rest/v1.0/projects/{project_id}/purchase_requestsAgents can propose
POSTPOST /rest/v1.0/projects/{project_id}/purchase_requests/{id}/transitions/rest/v1.0/projects/{project_id}/purchase_requests/{id}/transitionsDepends on action
POSTPOST /rest/v1.0/projects/{project_id}/expense_reports/rest/v1.0/projects/{project_id}/expense_reportsAgents can propose
POSTPOST /rest/v1.0/projects/{project_id}/expense_reports/{id}/transitions/rest/v1.0/projects/{project_id}/expense_reports/{id}/transitionsDepends on action
Scopesexpense_reports:readexpense_reports:write
osc operations search --module expense_reports lists all 8.
Prequalification4 GET6 writes
GETList prequalification submissions on a project/rest/v1.0/projects/{project_id}/prequalification/submissionsRead
GETList prequalification questionnaires available to a project/rest/v1.0/projects/{project_id}/prequalification/questionnairesRead
GETShow a prequalification submission/rest/v1.0/projects/{project_id}/prequalification/submissions/{id}Read
GETShow a prequalification questionnaire/rest/v1.0/projects/{project_id}/prequalification/questionnaires/{id}Read
POSTInvite a vendor to prequalify (create submission)/rest/v1.0/projects/{project_id}/prequalification/submissionsAgents can propose
Scopesprequalification:readprequalification:write
osc operations search --module prequalification lists all 10.
General ledger, AP, AR, banking, payroll, job cost, WIP and close.
GETList an automation's runs/rest/v1.0/companies/{company_id}/automations/{id}/runsRead
Scopesautomations:readautomations:write
osc operations search --module automations lists all 11.
Try it · real catalog
Search all 2,341 operations
Same rule as osc operations search: every word must appear in the module, summary, path or method. Runs in your browser on catalog revision 3eff1ae8265e; nothing is sent anywhere.
$ osc operations search --limit 10
Type to search. The catalog loads on first use (about 400 KB).
Under the hood
Four tools. One loop.
The connector does not dump 2,341 tool definitions into your model's context. It exposes a handful of tools and lets the model look things up:
os_search_operationscompact summaries, 20 at a time, with a cursor
os_describe_operationone operation's exact path, query and body schema
os_readruns a GET as you; responses are capped and marked when truncated
os_writeonly when writes are enabled; needs confirmed: true and never retries
Plus a usage resource that teaches the workflow and a project-review prompt. Returned record text is treated as data, never as instructions.
Agents get the same permissions as the person behind them, minus the ability to say yes. Security & privacy has the details.
01
Scopes
Narrow before you start
Every assistant token is scoped and expires. Read-only covers the 58 read scopes; read-and-write adds writes except webhooks and company administration.
Tokens last 7, 30 or 90 days, up to 10 active per person
Only a SHA-256 hash is stored; the token is shown once
Scopes narrow your access. They never add to it
02
Consent
You say yes, explicitly
Hosted connections open os.construction in your browser: you sign in, read what the assistant asks for, and choose Allow access or Deny.
Hosted grants last up to 14 days, then ask again
Hosted connections are read-only in the preview
The assistant works as you: your projects, your records
03
Propose-only
Agents draft. They do not decide.
Assistant tokens and hosted connections are agent credentials. Of 1,277 cataloged writes, 636 are proposals agents may make; 635 approve or administer and answer 403 for any agent.
Approve, certify, execute, post, pay, release, void, delete: refused
Setting status: "approved" through an update: refused
6 transition endpoints allow only their submit-style actions
04
Human approval
A person signs off in the app
The draft lands where your team already works. Someone with the approve permission reviews it in os.construction, exactly as if a colleague had drafted it.
"Agent credentials can create and propose changes but cannot approve…"
Approval rules and roles are unchanged
Built-in os.construction agents are Planned and will follow the same rule
05
Audit
Every call leaves a trail
Agent calls are recorded with the person they act for. Company admins review proposals and refused attempts under Company admin > API activity.
Tracked edits keep field-level before and after
Hosted calls return an invocation ID you can look up
Service accounts have their own call log
06
Revoke
One click, effective on the next call
Revoke a token in Account > AI assistants, or a hosted connection in Account > Connected apps. Changing your password or two-factor settings revokes all of your tokens.