MCP overview

There are two ways to reach Stratta. Both speak the Model Context Protocol, read the same corpus and count against the same quota; only the way you prove who you are differs.

The remote connector is the shorter path, and the only one that works in an agent running in the cloud, including ChatGPT, Codex cloud, and claude.ai. See Connect an agent from the browser.

The local server is still what you need for an agent that cannot open a browser, and for add_attachment, the one tool that reads a file from your disk. Ingestion is served to both: the pre-pass runs on the agent's machine, the writes go through whichever transport is connected.

Authentication#

Every tool call is authenticated. The remote connector uses a short-lived OAuth token; the local server uses an API key. Stratta resolves either credential to your user and organization and scopes all reads and writes accordingly. The organization is derived from the validated credential. It is never a parameter a client can set.

How your key is resolved#

The server looks for your key in this order and uses the first it finds:

1. STRATTA_API_KEY environment variable

   If set, it takes precedence. Best for ephemeral or CI-style usage.

2. ~/.stratta/config.json

   Written by npx -y @stratta/mcp login or the first-use prompt. Stored owner-only (mode 0600):

   { "apiKey": "sk_strt_xxx" }

3. First-use prompt (elicitation)

   If no key is found and your client supports it, the server asks for the key on the first tool call, validates it, and saves it to the config file.

Environment variables#

How your agent uses the tools#

Your agent is guided to navigate norms in a deterministic order so that every answer is backed by a real citation:

1. get_methodology

   The mandatory first call: it loads the persona, navigation rules, and citation format.

2. list_norms

   See which norms are available in your workspace.

3. get_toc → get_subtree

   Open the table of contents and drill into the relevant chapter.

4. get_section

   Read the full content of a section (text, formulas, tables, figures, cross-refs).

5. get_cross_refs → get_section

   Follow links to other norms and read those too, for compound questions.

6. get_figure

   Retrieve a diagram when the answer depends on it.

Use search_corpus when the norm itself is unknown, and search_in_norm whenever the section path is.

Resources and prompts#

Both transports serve four stratta:// resources, scoped to your workspace like the tools and not counted against your quota: stratta://norms (JSON), stratta://methodology (Markdown), stratta://norm/{code}/toc (table of contents to depth 2, code URL-encoded) and stratta://dossier/{id} (a project dossier as Markdown). A client that supports resources lets you pin one to a conversation; an agent can read one without a tool call.

The local server also offers three prompts, shown as slash commands where the client supports them: investigate-question, resume-dossier and check-value.

Usage tracking#

After each successful tool call, the server records a usage event (tool name, status, duration) against your key. This powers your dashboard stats and future billing. Tracking happens only after authentication and uses the key as proof of possession. Usage can't be attributed to someone else's key.

Tool families and simple mode#

Tools are grouped in four families: norms (always on: the 10 read tools, and remotely the citation cards, the norm reader and the profile), dossiers, sites and ingest (local only). The 59 local and 61 remote tools are the full mode, the default. Simple mode announces norms only: the agent loads fewer schemas and answers faster. Locally, use --simple or --toolsets; remotely, the Tools menu on the AI Agents page. See CLI and commands.

Tool reference#