Explore a small project below, then keep coding in Codex, Claude Code or Cursor. Connect SPEC//OS once and give your agent the project brief from the home page. The agent reads the context, manages specifications and records evidence here while it writes and tests code in its own environment.
This example shows the structure of a project. It is not a saved project, an applied specification or a test result. Viewing it needs no account and starts no AI operation.
INTENT
Let a person complete a task, keep that change after a reload, and keep each account's tasks separate.
Illustrative TypeScript web application
A task list with account-scoped persistence; no repository is connected.
EXAMPLE-RULE-001 · MUST — Preserve saved tasks when changing their completion state.
EXAMPLE SPECIFICATION
Keep task completion reliable and account data separate.
Proposed in this example · Not evaluated
EXAMPLE-AC-001.1
An identified reviewer confirms that completing a task and reloading preserves its completion state.
Verification method: identified human review
Proposed in this example · Not evaluated
Depends on EXAMPLE-REQ-001
EXAMPLE-AC-002.1
An identified reviewer confirms that completing a task in one account leaves another account's tasks unchanged.
Verification method: identified human review
Evidence: none · Not evaluated
Illustrative content only. No checks have been run and no evidence is attached.
What would happen next
Review task completion after a reload
EXAMPLE-REQ-001 → EXAMPLE-AC-001.1
This example has no recorded review for EXAMPLE-AC-001.1.
{
"schemaVersion": "specos.public-example/1",
"illustrative": true,
"readOnly": true,
"persisted": false,
"project": {
"id": "EXAMPLE-TASKBOARD",
"name": "Taskboard — public example",
"intent": "Let a person complete a task, keep that change after a reload, and keep each account's tasks separate.",
"context": {
"stack": "Illustrative TypeScript web application",
"architecture": "A task list with account-scoped persistence; no repository is connected.",
"constraints": [
"Preserve existing tasks",
"Keep account data separate"
]
},
"constitution": [
{
"id": "EXAMPLE-RULE-001",
"kind": "MUST",
"text": "Preserve saved tasks when changing their completion state."
}
]
},
"appliedSnapshot": null,
"serverRevision": null,
"spec": {
"objective": "Keep task completion reliable and account data separate.",
"requirements": [
{
"id": "EXAMPLE-REQ-001",
"title": "Keep completed tasks after a reload",
"source": "proposed",
"dependencies": [],
"criteria": [
{
"id": "EXAMPLE-AC-001.1",
"statement": "An identified reviewer confirms that completing a task and reloading preserves its completion state.",
"verifierId": "human_review"
}
]
},
{
"id": "EXAMPLE-REQ-002",
"title": "Keep each account's tasks separate",
"source": "proposed",
"dependencies": [
"EXAMPLE-REQ-001"
],
"criteria": [
{
"id": "EXAMPLE-AC-002.1",
"statement": "An identified reviewer confirms that completing a task in one account leaves another account's tasks unchanged.",
"verifierId": "human_review"
}
]
}
]
},
"evidence": [],
"verification": {
"status": "NOT_EVALUATED",
"reason": "Illustrative content only. No checks have been run and no evidence is attached."
},
"nextWork": [
{
"requirementId": "EXAMPLE-REQ-001",
"criterionId": "EXAMPLE-AC-001.1",
"title": "Review task completion after a reload",
"reason": "This example has no recorded review for EXAMPLE-AC-001.1."
}
]
}Compilation, chat and agent generation can consume credits. Their current credit cost depends on the selected company workspace and its configuration. Sign in and select that workspace to review the configured cost before approving a paid request.
There is no universal public credit price. Personal provider costs and monetary costs, including a USD conversion, are unavailable here. An unknown cost does not mean zero. Your coding agent or provider may charge separately.
The live catalogue marks audited MCP operations that consume no SPEC//OS credits with a credit cost of zero. Paid or unpriced operations report an unknown cost until the relevant scope can be checked. Reading this page or the illustrative example starts no paid request.
For a new project, save your idea on the home page. For an existing one, provide the GitHub repository and the next change. Open a recent project to recover its agent brief at any time.
Your agent uses its own authorized GitHub connection to create or clone repositories and prepare changes. The SPEC//OS GitHub App separately provides read access to the repositories you authorize. Adding a repository URL does not grant that access.
Choose a project/mcpStreamable HTTP with OAuth 2.1. The first connection opens a sign-in and consent page in your browser.
Assistants store the tool list on their side, at the moment the connector is created. Opening a new chat does not refresh it — the stale list follows your account into every new conversation.
This address answers without any sign-in and lists exactly what this server exposes right now:
/api/public/mcp-toolsIf the number there is higher than what your assistant shows, the assistant is the one out of date. Two ways to fix it, both requiring your sign-in:
/mcp inside an interactive session to authenticate:claude mcp add --transport http --scope user specos /mcpAfter reconnecting, open a fresh conversation and ask for the tool list — a session keeps the list it received when it started.
claude mcp add --transport http --scope user specos /mcp/mcp, then choose specos and authenticate. Your browser opens the SPEC//OS sign-in and a consent screen listing what the assistant may do./mcp again to confirm. The list looks like this: specos ✔ connected http (user)
tools: list_projects, project_state, compile_specification,
apply_revision, add_evidenceIf it shows “needs authentication”, run /mcp and authenticate again — the token lives on your computer and expires.
Settings → Connectors → Add custom connector, name it SPEC//OS and paste the endpoint. Claude opens the sign-in and consent page; after approving, the tools appear in the attachments menu of a conversation.
/mcpSettings → Connectors → Advanced → Developer mode, then Add custom connector and paste the endpoint. Approve the sign-in, then enable SPEC//OS in the composer for the chats that should use it.
/mcpAdd a remote MCP server (streamable HTTP) with this URL and complete the sign-in when the client opens it.
/mcpReturns the supported next steps for starting or continuing a project, including the work performed by your external coding agent.
Reads the complete current intake, constitution and applied snapshot before coding.
Creates a project in an explicit workspace, or saves reviewed repository context to an existing project.
Lists the GitHub App installation, authorized repositories and project bindings for a company.
Associates a project with a repository from the authorized installation. Requires the appropriate company role.
Uses the bounded repository reader to inspect the bound repository and propose project context.
Reading is immediate. Anything that writes or spends credit is stated below so the assistant can ask you first.
Lists every SPEC//OS project you can reach, in your own account and in the companies you belong to, most recently active first. The usual first call, to find the project to work on.
A compact status summary for one project. Use project_context for the full intake and applied specification.
Turns the current intent into a proposed specification, grounded in the project context and its decisions. It changes nothing on its own — the result is a draft for you to read. Quota, write permission and usage are checked before any cost, and repeating the same request does not charge twice.
Applies a draft as the project's next revision. The proposal token is required: if the intent, context, constitution or applied revision changed after the draft was produced, the request is refused instead of overwriting newer work, and the same token cannot be applied twice. The previous revision stays in history.
Records the result of a check you actually ran. The exact applied snapshot id is required; a newer revision is refused even if the capture time is recent. Quantitative checks require the observed value and matching unit. Recording evidence does not itself mean a criterion passed.