Docspeed

npx add-mcp 'https://api.docspeed.ai/mcp/'

Your agent. Docspeed document tools.

Read and structure documents, extract fields and tables, ask cited questions, compare document evidence, or build a reusable workflow with Docster inside the agent you already use.

Create an account Connect your agent

New accounts created directly with Docspeed start with 10 credits of API balance. Create an API key in your dashboard to use it. Paid top-ups are currently unavailable.

Native MCP connection + companion skill · No ZIP or Python required

Official MCP Registry record: ai.docspeed/docspeed · 0.1.0 (JSON). Install and connect below; a registry listing does not automatically install or enable tools in your client.

Connect now — no ZIP or Python required

Create your account and API key. Open your document project in a macOS or Linux terminal using Bash or zsh. Run this block, paste the key at the hidden prompt, and press Enter:

printf 'Docspeed API key (input hidden): '
read -r -s DOCSPEED_API_KEY
printf '\n'
export DOCSPEED_API_KEY

The key stays in this terminal's environment, outside command history and MCP configuration. Keep this terminal open for installation and launch; repeat the hidden prompt in a new terminal. Do not paste the key into chat or shared files. A separately launched desktop app does not automatically inherit this terminal's key; these steps launch the terminal clients.

Open a terminal in your document project. Choose your installed client and paste its command. It adds the native MCP connection and fetches the companion skill directly. Already customized this skill? Preserve it and use the ZIP installer's conflict checks.

Codex

codex mcp add docspeed --url https://api.docspeed.ai/mcp/ --bearer-token-env-var DOCSPEED_API_KEY && \
mkdir -p .agents/skills/docspeed-documents/agents .agents/skills/docspeed-documents/scripts && \
curl --fail --silent --show-error https://api.docspeed.ai/agent-tools/skills/docspeed-documents/SKILL.md -o .agents/skills/docspeed-documents/SKILL.md && \
curl --fail --silent --show-error https://api.docspeed.ai/agent-tools/skills/docspeed-documents/agents/openai.yaml -o .agents/skills/docspeed-documents/agents/openai.yaml && \
curl --fail --silent --show-error https://api.docspeed.ai/agent-tools/skills/docspeed-documents/scripts/upload.py -o .agents/skills/docspeed-documents/scripts/upload.py

Codex registers the MCP server in your user configuration and the skill in this project. From the same terminal, start a fresh session here:

codex -c mcp_servers.docspeed.tool_timeout_sec=1200

The timeout preserves normal permissions. For recurring use, set tool_timeout_sec = 1200 in the Docspeed server table below.

Claude Code

claude mcp add --transport http --scope project docspeed https://api.docspeed.ai/mcp/ --header 'Authorization: Bearer ${DOCSPEED_API_KEY}' && \
mkdir -p .claude/skills/docspeed-documents/scripts && \
curl --fail --silent --show-error https://api.docspeed.ai/agent-tools/skills/docspeed-documents/SKILL.md -o .claude/skills/docspeed-documents/SKILL.md && \
curl --fail --silent --show-error https://api.docspeed.ai/agent-tools/skills/docspeed-documents/scripts/upload.py -o .claude/skills/docspeed-documents/scripts/upload.py

Claude registers both in this project. Keep the server name and URL before --header, and keep the single quotes so the saved configuration contains the variable name, never the key. From the same terminal, launch:

MCP_TOOL_TIMEOUT=1200000 claude

If claude mcp get docspeed shows Pending approval, start the interactive session above, accept the workspace trust prompt for your project, and review and approve the docspeed project MCP server when asked. Registration alone does not approve the connection.

Your first grounded result

Inside the new client session, enter /mcp (not in your terminal shell). Follow the client's normal trust and tool permission prompts: review the connection, then allow the requested document operation when you intend to send that document. Confirm Docspeed is connected and the docspeed-documents skill is available. Attach an authorized invoice and ask:

“Extract the invoice number, supplier, currency, tax, total, and every line item. Show the supporting page and source evidence for each value. Mark values absent from the document as missing.”

Check each returned value against the original and open its source evidence. Uploading sends the document to Docspeed and normal processing charges apply. A connected server proves discovery, not document accuracy.

If MCP is still connecting, the skill retries discovery once. If tools remain unavailable, it explicitly offers native/local reading for you to choose. Ask provides source boxes that make answers easier to check in scans and long documents.

Read the skill before installing. Commands follow the Codex MCP documentation and Claude MCP documentation.

Upload a local document in one command

With DOCSPEED_API_KEY set in the same terminal, run the installed helper on a document you authorize sending to Docspeed. Python 3.10+ is needed for this optional helper, with no extra packages:

# Codex project:
python3 .agents/skills/docspeed-documents/scripts/upload.py "/absolute/path/to/document.pdf"

# Claude Code project:
python3 .claude/skills/docspeed-documents/scripts/upload.py "/absolute/path/to/document.pdf"

The helper streams multipart bytes directly to POST https://api.docspeed.ai/api/v1/upload using the existing files form field. It keeps document bytes out of your agent conversation and prints only:

{"file_id":"your-returned-file-id","filename":"document.pdf"}

Give that reference to your agent with your document question. Use file_id in each tool's actual file_id or file_ids parameter; Docster samples also use filename. Reuse it across calls. The upload and MCP connection must belong to the same account.

The helper accepts one nonempty supported PDF/image up to 50 MiB; the API validates it. It reads the key only from the environment, requires verified HTTPS, and follows no redirects or automatic retries. If a connection fails after sending bytes, the upload outcome can be unknown; do not blindly retry. No URL publishing or new credential flow is needed. Uploading does not itself run extraction.

Use direct upload for anything larger than 4 KiB. For a tiny files_upload input only, this portable Python/stdin command works on macOS and Linux and refuses larger files:

python3 -c 'import base64,sys; data=sys.stdin.buffer.read(4097); len(data)<=4096 or sys.exit("Use direct upload for files larger than 4 KiB"); print(base64.b64encode(data).decode("ascii"))' < "/absolute/path/to/tiny.pdf"

Pass that tiny result as content_base64 with the original filename. Never print a larger document's base64 into the conversation.

Inspect a result before integrating

Open the original synthetic invoice beside its recorded extraction. Select a field or line-item cell to see the exact source region returned with that value, then download the PDF and result JSON.

Explore the source-review example

One recorded Codex run: 11 invoice fields and 15 line-item values were checked against the source. This is a historical synthetic test, not a live demo, a customer result or a general accuracy guarantee. No API key or processing charge is needed to inspect it.

Or inspect a recorded Q&A and follow-up with source citations. This synthetic invoice example shows two saved Ask calls from an internal MCP check, with the returned answers and source areas. It is not a recorded agent conversation or a live processing demo.

Compare two synthetic agreements with cited Q&A. This recorded direct API run shows a notice-period comparison and its source boxes. The initial fee answer omits the payer and due date; both limitations are visible alongside the original passages.

1. Inspect the evidence. Compare a returned value with the text inside its source box. The example preserves printed zero and numeric types.
2. Connect the tools. Follow the setup below using your own account and the package's full capability guide.
3. Try your document task. Use the direct upload command for an authorized PDF/image and give its returned file ID to your agent. Tiny inputs can also use files_upload. The hosted MCP cannot read a local path or ingest a URL on its own. Normal processing charges apply.

Choose the operation your application needs: OCR or Parse for document content; Classify and Split for packets; Layout, Tables, Charts or Equations for page elements; Schema Auto and Extract for fields; Ask for cited answers; Compare for corresponding document values. Synthesize uses its separate API and is not exposed in the current MCP catalog; fidelity limits remain under validation. Docster adds reusable workflows through the authenticated interface described below. The download explains the supported paths and limitations.

MCP compare returns a compact verdict, items and source evidence by default. Set include_extractions: true for full per-document extractions. Both modes keep source page geometry and unchanged grounding regions; existing HTTP API clients keep the full response by default.

ZIP installer fallback

Download the integration ZIP. Unpack the download and open a terminal inside docspeed-agent-tools. Use Python 3.10 or later. On Python 3.10, install its TOML dependency with python3 -m pip install -r requirements.txt; Python 3.11+ uses its standard library.

Set your API key with the hidden prompt above, then run the helper and launch your client from that same terminal. Keep the key out of chat and shared files.

1. Check the connection

python3 scripts/setup.py check --catalog-out /tmp/docspeed-tools.json

The check authenticates, initializes MCP, and lists the available tools. It does not upload a document or run extraction.

2. Install for your client

python3 scripts/setup.py install --client codex --project /absolute/path/to/your/project

# For Claude Code:
python3 scripts/setup.py install --client claude --project /absolute/path/to/your/project

The installer adds the project connection and skill, preserves unrelated settings, and stores only the credential variable name. Conflicting existing Docspeed settings require manual review.

3. Start a fresh session

Open the project in your agent, follow its normal trust and permission controls, and inspect /mcp. The connection must be active before tools can be selected. Installing the package does not inject tools into an existing conversation.

Optional Codex direct MCP configuration

For Codex clients with Code Mode, add --codex-direct-mcp to the installer command to expose the DocSpeed namespace directly. This configuration was used in a recorded native MCP call with Codex CLI 0.159.2; it is not a universal requirement or a guarantee of selection. The full catalog remains available and your model, trust, and approval settings remain unchanged. Review existing custom Code Mode settings before merging the included config/codex-direct-mcp.toml snippet.

Claude Code project servers also need the client's normal MCP approval. A healthy MCP connection is separate from access to a Claude model.

Ask for the result you need

“Extract the supplier, invoice number, tax, total, and every line item from this PDF. Include evidence for each value.”

“Across these policy PDFs, which exclusions apply to flood damage? Cite the relevant pages.”

“Build a reusable invoice, purchase order, and receipt reconciliation workflow with a review screen and an API endpoint.”

The companion skill helps the agent recognize these tasks even when you do not name Docspeed. Selection still depends on the request, available tools, and permissions. Explicit provider choices and local-only requirements take precedence.

Build with Docster

Docster workflow builds are available to authenticated Docspeed accounts using a standard Docspeed API key. Normal usage charges and workflow validation apply.

Run python3 scripts/setup.py check --require-docster to require the complete build, status, revise, and invoke interface. If those tools are missing, the connected deployment cannot accept a Docster handoff.

docster_build starts with clarification, retains an attempt_id, then builds the requested workflow. Check docster_status and its validation evidence; a solution is ready only when its status is ready. Use docster_revise to update it and docster_invoke to run it on new uploaded documents.

Document tools run synchronously. Legacy agent-definition and job APIs are not exposed through this MCP catalog. Building a private workflow does not publish it or authorize sending results to other people.

Manual connection

Streamable HTTP endpoint: https://api.docspeed.ai/mcp/. Authentication uses a Docspeed API key as a bearer token. Keep the trailing slash.

Codex configuration

[mcp_servers.docspeed]
url = "https://api.docspeed.ai/mcp/"
bearer_token_env_var = "DOCSPEED_API_KEY"
startup_timeout_sec = 30
tool_timeout_sec = 1200

Add this to your project .codex/config.toml or user configuration. The download also contains Claude Code configuration, portable plugin manifests, the companion skill, and a routing evaluation runner.

Inputs and evidence

Production supports PDF, PNG, JPEG/JPG, TIFF/TIF, WebP, BMP, and GIF. Multi-page TIFF is page-indexed. Native Word/Excel (DOCX/XLSX) workflows are available through separate beta APIs by contacting Docspeed.

Uploading sends the selected documents to Docspeed; use only files authorized for the workflow. Processing and builds use your normal API quota and billing. Grounded results preserve their source regions; missing values and processing failures must remain explicit.

A successful connection check proves discovery, not document accuracy or workflow completion. The included selection evaluation measures routing decisions separately from real tool execution and document quality.

Full guide included in the download · Codex MCP reference · Claude Code MCP reference