skill
docspeed-agent-tools
Docspeed document tools and skill for coding agents, with source-reviewed recorded examples
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/docspeed-ai/docspeed-agent-tools itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 28, 2026Docspeed for coding agents
Connect Codex or Claude Code to Docspeed's hosted document tools and install the companion docspeed-documents skill. Use your existing agent to extract fields, read documents, answer cited questions, compare document values, or build a reusable Docster workflow.
This MIT-licensed repository contains client integration assets, not the Docspeed server or a local/offline document engine. Public setup guide · API documentation · Recorded examples
Connect with an installed client
The public guide now provides direct setup without downloading this repository or installing Python. Use a macOS or Linux terminal with Bash or zsh, in your document project. Choose this route or the repository installer below. If you already customized the companion skill, use the repository installer's conflict checks instead of overwriting it with these download commands.
With your account API key ready, read it at the hidden prompt in the same terminal that will install and launch the client:
printf 'Docspeed API key (input hidden): '
read -r -s DOCSPEED_API_KEY
printf '\n'
export DOCSPEED_API_KEY
The key stays in the terminal environment; it is not saved in command history or client configuration. A separately launched desktop app does not inherit it automatically.
For Codex, add the native connection and download the companion skill and upload helper:
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
This registers the server in your user configuration and the skill in the current project. Start a fresh session from the same terminal:
codex -c mcp_servers.docspeed.tool_timeout_sec=1200
For Claude Code, register the connection and skill in the current project:
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
Keep the single quotes around the header so the configuration stores the environment variable name, not its value. Launch from the same terminal:
MCP_TOOL_TIMEOUT=1200000 claude
Follow the client's normal project trust and MCP approval prompts, then inspect /mcp inside the session. These commands configure the connection; they do not bypass permission prompts, upload a document, prove model entitlement, or guarantee tool selection. Python is only needed for the optional direct-upload helper and repository installer described below.
Install from this repository
You need Python 3.10 or later, an installed Codex or Claude Code client, and a Docspeed API key from your account dashboard. Make DOCSPEED_API_KEY available to the process that launches your client. Do not paste the key into chat, commit it, or pass it as a command argument. A desktop client must receive the environment variable too.
Download or clone this repository, then open a terminal in its root directory. Python 3.11+ needs no additional package. On Python 3.10 only:
python3 -m pip install -r requirements.txt
Install the connection and skill into an existing project, replacing the example path:
python3 scripts/setup.py install --client codex --project /absolute/path/to/your/project
For Claude Code, use --client claude. This local install does not upload or process documents. It preserves other configured servers and client approval settings; conflicting Docspeed configuration or a modified skill stops the install for manual review.
Start a new agent session in that project, complete the client's normal project/MCP trust prompts, and inspect /mcp. Installation alone does not inject tools into an already-running conversation or guarantee tool selection.
With your key set, check the real connection:
python3 scripts/setup.py check
The check initializes MCP and lists its tools; it makes no document-processing or model call. Success proves connection and discovery of the checker's required core tools, not every operation, sufficient balance, output quality, or completed workflows. For the four Docster tools specifically:
python3 scripts/setup.py check --require-docster
The canonical MCP endpoint is https://api.docspeed.ai/mcp/, including the trailing slash. The checker rejects redirects rather than forwarding credentials to another origin. It reports errors explicitly and does not switch providers.
Try a document task
Use a document you are authorized to send to Docspeed:
Extract the supplier, invoice number, taxes, and all line items from this PDF. Include source evidence and identify missing values.
Or ask:
Which cancellation terms apply across these contract PDFs? Cite the supporting pages.
Choose the smallest applicable operation from the actual connected tool catalog. The skill covers OCR, Parse, Layout, extraction, document questions, tables, charts, equations, classification, splitting, comparison, and reusable workflows when their tools are exposed. schema_auto can propose a schema when it is available; never invent a missing capability.
For reusable work, ask for a Docster workflow. Builds are available through the authenticated interface and require the normal quota, billing, and validation controls. Begin with clarification, retain the returned attempt ID, and report completion only when the workflow is ready.
Production inputs are PDF and PNG/JPEG/TIFF/WebP/BMP/GIF images. Native DOCX/XLSX processing has a separate beta API path. The remote MCP cannot read a local filesystem path or ingest a URL.
For an authorized local file, use the shipped Python 3.10+ helper from the installed skill directory. It needs no extra package:
# 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"
It streams one supported, nonempty PDF/image of at most 50 MiB directly to the hosted upload endpoint using DOCSPEED_API_KEY from the environment. It prints only the reusable file_id and submitted filename. Keep the upload and MCP calls on the same account. Uploading does not run extraction. The helper verifies HTTPS and follows no redirects or automatic retries; an interrupted upload can have an unknown outcome.
Keep document bytes out of the agent conversation. files_upload remains an option for tiny inputs of at most 4 KiB; use direct upload for larger files. Reuse the returned file reference for subsequent operations. The skill contains the exact supported workflow and failure rules.
Synthesize is API-only: it is not exposed in the current MCP catalog. Its separate public API is documented in the Synthesize guide. Reference fidelity remains under validation. Do not substitute a workflow build or claim generated content is evidence about a real source.
Other installation choices
- Manual Codex configuration: merge
config/codex.tomlinto your chosen configuration and copyskills/docspeed-documents/into the chosen skill directory. Preserve other settings. - Optional Codex direct namespace mode: use
--codex-direct-mcpwith the project installer, or reviewconfig/codex-direct-mcp.toml. This is opt-in, preserves approval policy, and does not establish support or selection for every Codex version. - Local Claude Code plugin: with your key in the environment, start
claude --plugin-dir /absolute/path/to/this/repository. Use this or the project installer to avoid duplicate registrations. - The Claude catalog now pins the matching ZIP checksum
8435e968cb9fec1de526d623f608fe54cc0de682a73cbdfe9b5a87545ac3a21b, verified on 12 October 2026. This repository and the direct setup commands above do not use that catalog.
plugin.json/mcp.json also provide portable Agent Plugins metadata. Compatible host support and directory submission are separate. The official MCP Registry record describes the hosted server; a listing does not install or enable it in your agent.
Privacy, cost, and result boundaries
Uploading sends the selected document to Docspeed. Respect local-only requirements and the user's chosen provider. Do not upload private documents without authorization. Normal Docspeed charges apply to processing and workflow calls; use your account's available balance. Paid top-ups are currently unavailable. Client/model subscription or entitlement is separate from Docspeed access.
Inspect returned citations, page geometry, absent values and errors against the source. A successful tool response is not an accuracy guarantee. Preserve explicit failures rather than silently changing OCR engines or providers. Recorded examples are bounded synthetic evaluations, not customer outcomes or universal quality claims.
Read How to check document answers for completeness for a source-backed review of the three recorded examples and their limits.
Package contents and provenance
docs/document-answer-completeness.md: reviewed write-up of historical synthetic evaluations.skills/docspeed-documents/: portable skill, Codex metadata and direct-upload helper.config/,.mcp.json,mcp.json: environment-based or host-managed key configuration.scripts/setup.py: project installer and bounded MCP connection check..claude-plugin/plugin.json,plugin.json: plugin manifests.evals/,scripts/eval_selection.py: optional metadata-selection evaluation. Running it invokes your installed model client and can consume its usage; it does not measure live document quality or automatic skill discovery.
The client configuration, skill and installer assets are exact copies of the published ZIP identified in PROVENANCE.json; the Claude catalog is copied separately from its published URL. This standalone README retains the public repository's direct setup and launch instructions, with licensing and checksum updates; it is not a copy of the ZIP's monorepo-oriented README. PROVENANCE.json identifies both README sources and their hashes, and MANIFEST.json records the public file set and current hashes. No server code, account credentials, customer documents or private execution records are included.
License scope
The companion integration package is licensed under the MIT License. This grant covers the distributed installer, skill, helper, manifests and package documentation. It does not license the private Docspeed core, hosted API/MCP/Docster implementation, customer documents or service access. Hosted usage remains subject to the account's terms, balance and quotas. See LICENSE_STATUS.md for the scope.