For AI Agents

Connect coding agents to ZeroDrift’s machine-readable documentation, OpenAPI specification, and docs search MCP server.

Use these public, read-only surfaces to discover and query the ZeroDrift documentation while you code. They are generated and hosted by Mintlify from the same source as this site.

Machine-readable documentation

Surface	URL	Use it for
Documentation index	https://docs.zerodrift.com/llms.txt	Discover every published page and its one-line description.
Full documentation	https://docs.zerodrift.com/llms-full.txt	Load the complete documentation corpus into a single context.
Markdown pages	Append .md to any page URL, such as /quickstart.md	Read one page as Markdown without site navigation or presentation markup.
OpenAPI specification	https://docs.zerodrift.com/openapi.json	Inspect the API contract or generate a client.
Documentation MCP	https://docs.zerodrift.com/mcp	Search and retrieve published ZeroDrift documentation from an MCP client.

The documentation MCP is distinct from MCP Connectors. The documentation MCP lets your coding agent read ZeroDrift’s public docs. MCP Connectors let ZeroDrift import your policy documents from connected services.

Add ZeroDrift guidance to your repository

Paste the following into your repository’s AGENTS.md or CLAUDE.md:

## ZeroDrift integration

ZeroDrift Enforcement API. Base URL: https://api.zerodrift.ai. Authenticate with the `x-api-key` header using the `ZERODRIFT_API_KEY` environment variable. Never expose the key in source control, logs, client-side code, or prompts.

Enforce text: `POST /api/v3/content/validate` with `{content, model_engine: "anchor_3_0_mini"|"anchor_3_0"|"anchor_3_0_max", mode: "sync"|"async", validation_scope?: {rulepacks?, imports?, rules?, managed_only?}, metadata?}`. Default `model_engine` is Flagship (`anchor_3_0`; older notes called this Base). Mini is detection-only, runs unscoped against its frozen rulebook, and rejects `validation_scope`. A scope-less Flagship call stays on Flagship and checks active managed rules only; activated custom rules are not evaluated. Ask for `anchor_3_0_max` explicitly to enforce activated custom rules. Max returns per-line violations without a rewrite; `fix.suggested_text` is the rule's own fix note when present and otherwise empty. A Flagship request that explicitly names a `cust_*` rule is upgraded to Max (`engine_upgrade_reason: "custom_rules"`). See [Choose a model](/models). Quality and latency: https://zerodrift.com/model/anchor-3-benchmark

Sync mode returns the result inline when processing completes within the inline budget; otherwise it falls back to async mode. Custom rules need `anchor_3_0_max` for an inline sync verdict. Async mode returns `202` with `{job_id, poll.url}`; poll `GET /api/v3/jobs/{job_id}` until `status` is `completed` or `failed`.

Read the enforcement decision from `result.overall_status`: `approved` = Pass (deliver); `needs_changes` = Fix (apply `fix.suggested_text` only when `fix.action` is `replace`); `do not send` = Block (deliver a fallback, never the original). Treat `status` only as job lifecycle state. Escalate when `suggested_text` is empty, `action` is not `replace`, or your policy requires review. Public usage fields are `result.units` and `result.cost`, not raw token counts.

Rulepacks: `GET /api/rulepacks/`; use the returned `id` in `validation_scope.rulepacks`. Rules: `GET /api/policies/rules`; use `rule_id` in `validation_scope.rules` (custom IDs are prefixed `cust_`). Policies: `POST /api/policies/import`, poll the import, activate its rules, then use the import ID in `validation_scope.imports`.

Activity log (30 days): `GET /api/activities`. Rate limit: 10 requests per second with a burst of 50; back off on `429`.

Docs index: https://docs.zerodrift.com/llms.txt
Full docs: https://docs.zerodrift.com/llms-full.txt
OpenAPI spec: https://docs.zerodrift.com/openapi.json
Docs MCP: https://docs.zerodrift.com/mcp

Start an integration

Quickstart

Follow the current onboarding flow, then ask your coding agent: “Add ZeroDrift enforcement to our support bot. Check each generated reply before sending it, apply safe fixes, block unsafe replies, and log escalations.”