# SDD Vault is live: an MCP server so your AI agents work from one spec

> Source: <https://dev.to/eduin_garciacordero_c315/sdd-vault-is-live-an-mcp-server-so-your-ai-agents-work-from-one-spec-4jm2>
> Published: 2026-11-20 19:06:55+00:00

**Key takeaway:** SDD Vault exposes MCP tools including `claim_task`, which refuses a claim on a taken task and names the holder, so parallel agents can share one task list without duplicating work.

Three weeks ago a Claude Code session rewrote our backend from Rust to TypeScript. The spec it worked from had 62 tasks, and the agent could not edit it. That is the product, and today it is open to everyone.

A hosted store for Spec Driven Development documents (constitution, specs, plans, tasks, research, validation) with an MCP server in front. Your agents read and write the documents through MCP instead of the file system. People read, answer open questions and approve in the browser, without git.

Git is great for code. Specs that many agents work from at once need a server that knows what a task, a requirement and a principle are:

|  | Files in git | SDD Vault | 
|---|---|---|
| Two agents, one task list | Both start the same task on two branches | Each agent claims a free task with `claim_task` ; a claim on a taken one is refused with the holder's name, and the app shows who is on what | 
| "Which FRs have no task?" | Every agent parses every file | One call, answered from an index | 
| History | `git blame` names the committer, not the agent; rebases rewrite it | Every write is a revision with the person, the agent and the time | 
| The constitution | One copy per repo, anyone edits it | Written once at the organization, inherited by every project, writable by owners only | 
| Approval | A word in a header | A recorded state transition with a name and a time; open markers block it | 

Claude Code:

```
claude mcp add --transport http sddvault https://{org}.mcp.sddvault.com/mcp
```

Cursor, `.cursor/mcp.json`:

```
{ "mcpServers": { "sddvault": { "url": "https://{org}.mcp.sddvault.com/mcp" } } }
```

The agent opens the browser, you approve it once, and it never sees a password. The tools it gets: `get_agent_guide`, `project_status`, `list_projects`, `list_specs`, `get_active_spec`, `get_document`, `put_document`, `create_spec`, `set_spec_status`, `list_tasks`, `claim_task`, `release_task`, `mark_task_done`, `list_markers`, `list_revisions`, `get_revision`, `get_constitution`, `put_constitution` (owners), and the template tools.

```
curl -fsSL https://get.sddvault.com/install.sh | sh
sddvault login --org acme
sddvault import acme/demo . --dry-run
sddvault import acme/demo .
```

SDD, GitHub Spec Kit and Kiro layouts import. `sddvault export` writes the same files back, byte for byte.

The full case study is here: [https://dev.to/eduin_garciacordero_c315/how-we-rewrote-our-backend-with-ai-agents-working-from-specs-2d84](https://dev.to/eduin_garciacordero_c315/how-we-rewrote-our-backend-with-ai-agents-working-from-specs-2d84). The short version:

`make test` from 46 minutes to 7, every performance threshold kept.
Free: 3 people, 1 project, 10,000 MCP calls a month, 30-day history, no card. Pro ($10 per user a month) and Team ($15 per seat a month) for unlimited projects, history and organizations with roles, approvals and audit. Readers and reviewers are free on every plan.

If you run agents in parallel, tell me in the comments what breaks first for you. That is the list I am working from.
