Accie OCR

MCP

The Accie OCR MCP server is a thin stdio wrapper around the same HTTP API at https://ocr.accie.ai. It does not add a second product surface. File bytes are not sent through MCP — you PUT them to the upload_url the API returns.

Server name accie-ocr. Tools call the REST API with ACCIE_API_KEY or ACCIE_BEARER_TOKEN.

Run it

Install and run:

pip install mcp httpx
ACCIE_API_URL=https://ocr.accie.ai ACCIE_API_KEY=accie_sk_... \
  python mcp/accie_ocr_server.py

Cursor MCP config (do not commit keys):

{
  "mcpServers": {
    "accie-ocr": {
      "command": "python",
      "args": ["/absolute/path/to/accie_ocr/mcp/accie_ocr_server.py"],
      "env": {
        "ACCIE_API_URL": "https://ocr.accie.ai",
        "ACCIE_API_KEY": "<from POST /api/v1/api-keys>"
      }
    }
  }
}

Authentication (env only)

Variable	Role
`ACCIE_API_URL`	API base. Production: `https://ocr.accie.ai`.
`ACCIE_API_KEY`	Preferred. Sent as `Authorization: Bearer …`.
`ACCIE_BEARER_TOKEN`	Used only if the API key env is empty (session or Accie JWT).

If neither secret is set, the server raises: Set ACCIE_API_KEY (preferred) or ACCIE_BEARER_TOKEN — never hard-code secrets. Every tool goes through that header helper, including tools that wrap public HTTP routes such as list_export_formats.

Tools

Each tool returns a JSON string (pretty-printed on success). HTTP calls use httpx with a 60s timeout.

Tool	HTTP	Notes from the server docstring / code
`list_ocr_kinds`	`GET /api/v1/ocr-kinds`	Five primary OCR kinds with org enablement. Unauthenticated call: HTTP 401 `missing_bearer_token`.
`create_ocr_job`	`POST /api/v1/jobs`	Required `filename`. Defaults: `document_class=invoice`, `estimated_pages=1`, `content_type=application/pdf`, `prefer_gpu=false`. Optional: `enhance`, `id_type_hint`, `script_profile`. Legacy `indic` maps to handwritten + `script_profile=indic`. ID jobs set `options.enhance` default `auto` and `validate=true`. Returns `upload_url` — PUT the file yourself, then poll `get_job_status`.
`get_job_status`	`GET /api/v1/jobs/{job_id}`	Status, page_count, document_class, `documents[]` for ID, error_message if any.
`export_job`	`POST /api/v1/jobs/{job_id}/exports/{fmt}`	Default `fmt=json`. Docstring: short-lived `download_url`, pages are not re-charged. Formats follow the live export matrix.
`usage_summary`	`GET /api/v1/usage`	Pages by OCR kind, feature events, plan entitlements. Auth required (not captured here).
`list_export_formats`	`GET /api/v1/export-formats`	Public HTTP route. Body is on the REST API page. MCP still requires an API key in env.
`validate_indian_id`	`POST /api/v1/id/validate`	Soft-validate without a job. Types: pan, aadhaar, dl, voter, passport, ration, pension, pran. Docstring: never government-verified. Aadhaar full numbers in the request may be masked in the response.
`list_id_schemas`	`GET /api/v1/id/schemas`	Public HTTP route. Response includes `government_verified: false` and disclaimer “Soft validation only — format/checksum/heuristic checks. Not UIDAI, NSDL, Parivahan, NVSP, or DigiLocker authentication.” MCP still requires an API key in env.

On HTTP 4xx/5xx, create_ocr_job, get_job_status, export_job, validate_indian_id, and list_id_schemas return a JSON object with error: true, status, and body instead of raising. list_ocr_kinds, usage_summary, and list_export_formats call raise_for_status().

Try it — the two public tools, without an MCP client

list_export_formats and list_id_schemas just call public HTTP routes under the hood. Run the same request an MCP agent would make:

Try it live — no key required

Every other tool — create_ocr_job, get_job_status, export_job, usage_summary, validate_indian_id, list_ocr_kinds — needs a real API key, since it acts on an organisation's data. See the REST API page for the full authenticated flow.

What MCP is not

It is not a second API with different auth. Same bearer / API key as REST.
It does not upload file bytes. Put bytes to upload_url, then poll get_job_status.