14-day free trial — no credit card requiredStart trial

ClawQL — the Agentic Gateway for Auditable Production AI

The Foundational Platform agents call into — not an agent framework. Land with OpenAI-compatible inference and native /mcp, then expand into memory, Dedicated Virtual Gateway governance, and Edge Gateways across your fleet. One closed loop: recall, search, execute, and ingest — with specs server-side, not in your prompt.

Start free trial Self-host free

$

npm install -g clawql-mcp

Workflows in practice

What a real ClawQL session looks like.

Published case studies document the same pattern: agents call MCP tools in sequence — recall context, discover operations with search, act with execute, then persist outcomes with memory_ingest. These feeds show the chronology without pasting specs into the chat.

Cloudflare docs MCP case study

Ship docs.clawql.com

Agent intent
Deploy the Next.js docs site to docs.clawql.com on Cloudflare Workers and verify custom-domain routing.
Recall prior context
memory_recall()
Query the vault for “Cloudflare docs deploy”, Workers 1101, and OpenNext fs failures — skip dead ends from earlier sessions.
Discover control-plane ops
search()
Rank Cloudflare REST operations for Workers custom domains and zone bindings — no multi‑MB OpenAPI in the prompt.
Attach hostname
execute()
workers.domains.update — bind docs.clawql.com to the clawql-docs Worker with account_id, zone_id, and service name.
Deploy and verify
OpenNext build via Wrangler; HTTP 200 on /; wrangler tail clean after fixing fs.readdir and MDX prerender crashes.
Persist runbook
memory_ingest()
Append deploy decisions and fixes to the vault with append: true and wikilinks — durable recall for the next incident.

Read case study

OpenClaw + memory_recall case study (June 2026)

OpenClaw recalls a Cursor roadmap

Months earlier — different agent, different product
In Cursor, memory_ingest writes roadmap priorities, vendor analysis, and K8s debugging notes to the vault. OpenClaw is not installed yet.
Fresh OpenClaw thread — zero pasted history
A new gateway agent comes online: “fresh slate, no memories.” Different product, different session, no chat scrollback from April.
Recall over Streamable HTTP
memory_recall()
User directs clawql__memory_recall with query “comprehensive project summary”, limit 5 — return vault JSON paths and snippets only, no guessing.
Ranked notes from long-ago sessions
Vault returns Memory/_INDEX_ClawQL.md, openclaw-mac-mini-clawql-mcp-setup.md, and docker-desktop-kubernetes-mcp-kyverno-cosign-session.md — each with path, score, and snippet from the tool response.
Portable institutional memory
Roadmap and runbook context written before OpenClaw existed surfaces in a new gateway chat — same CLAWQL_OBSIDIAN_VAULT_PATH, no copy-paste.

Read case study

Vault memory + GitHub session case study

Vault ingest → GitHub → audit

Ground on prior work
memory_recall()
Recall roadmap notes, vendor analysis, and open prioritization before triaging the backlog.
Capture decisions
memory_ingest()
Ingest structured insights with wikilinks — durable narrative, not ephemeral session cache.
Find GitHub issue APIs
search()
Discover issues/create and related operations from the bundled GitHub provider.
Open tracking issues
execute()
Create canonical GitHub issues for audit (#89), Helm wiring, and duplicate consolidation.
Log MCP events
audit()
Append structured breadcrumbs during the ship — correlate operator events with vault ingests later.
Close the loop
memory_ingest()
Append session outcomes so the next chat recalls what shipped and why.

Read case study

Context compression

~862× compression — 7,174 endpoints indexed, 62 surfaced per workflow.

The answer to “won't this just fill up my context?” Offline benchmark across Google Cloud, Cloudflare, and Jira: stand up GKE, configure Cloudflare DNS, and draft a Jira rollout — with planning tokens dropping from ~10.2M to ~12K.

Read the benchmark

~862×

Compression ratio on the three-provider benchmark — measured with ceil(bytes ÷ 4) token estimates.

7,174

API operations indexed across Google Cloud (4,141), Cloudflare (2,697), and Jira (336) — specs stay in the MCP server.

62

Unique operations returned by search across 14 workflow steps — the candidates an agent actually needs.

~10.2M → ~12K

Planning-context tokens: naive full-spec paste vs compact workflow artifact (~99.9% reduction).

How it fits together

Memory, APIs, and documents in one closed loop.

A typical session: an agent recalls prior vault notes about a deployment, searches Cloudflare and GCP for the right DNS and cluster operations, executes them with validated args, then ingests the outcome so the next session — in Cursor, OpenClaw, or your cluster — picks up where this one left off. Document workflows follow the same pattern through the IDP pipeline.

MCP tools

Every tool has a job — and a clear boundary.

Core discovery and execution are always on. Memory, documents, automation, and the IDP pipeline opt in when you need them — same surface in Cursor, OpenClaw, or your Kubernetes cluster.

ClawQL Core

Always on — no opt-out

search

Rank thousands of API operations by natural-language intent. Specs stay server-side — agents get operation IDs and parameter hints, not megabytes of YAML.

execute

Run one operation with validated args over REST, GraphQL, or gRPC. Lean responses keep tool results out of your context budget.

audit

In-process ring buffer of structured events — append, list, clear. Pair with memory_ingest when you need a durable operator trail.

cache

Ephemeral LRU scratch space for the active session. Use memory_* when data must survive restarts.

Memory & knowledge

Default on — opt out with CLAWQL_ENABLE_MEMORY=0

memory_ingest

Write Obsidian Markdown with wikilinks and structured insights. Decisions, runbooks, and postmortems persist across chat sessions.

memory_recall

Keyword search, wikilink graph hops, optional vector KNN. Agents pull institutional knowledge instead of re-asking the same questions.

ingest_external_knowledge

Bulk Markdown or HTTPS URLs into the vault — ideal for importing docs, threads, and vendor analyses.

knowledge_search_onyx

Hybrid enterprise search over your Onyx index. Ground agent answers in indexed documents, not model guesswork.

Automation & ops

Opt in via CLAWQL_ENABLE_* flags

schedule

Persisted synthetic HTTP checks with assertion-based pass/fail and run history — uptime without a separate monitoring stack.

notify

Slack chat.postMessage wrapper for workflow milestones. Pipe IDP completion summaries straight to your team channel.

sandbox_exec

Isolated code snippets via Seatbelt, Docker, or Kata — validate transforms without giving agents a raw shell.

workflow / argocd

Submit Argo Workflows and observe Argo CD Applications from the same MCP surface your agents already use.

IDP document pipeline

Eight bundled vendors, one compose path

run_idp_pipeline

Automated multi-hop pipeline: Nextcloud intake → Docling/Tika → Gotenberg → Stirling → archive → Onyx → Coneshare VDR.

classify_document

Route documents by type before extraction — W-2s, invoices, and contracts hit the right downstream handlers.

extract_document

Schema-grounded LangExtract fields with char_interval provenance — grounded extraction, not free-form hallucination.

hitl_enqueue_label_studio

Push low-confidence extractions to Label Studio for human review, then ingest decisions back into the vault.

Intelligent document processing

Eight vendors. One agent-composable pipeline.

Drop a file in Nextcloud; agents orchestrate layout parsing, PDF normalization, PII redaction, archival, hybrid search, and secure sharing — via search → execute or automated run_idp_pipeline. Self-host the full stack with Helm; managed accounts run the ClawQL Archive Layer (Nextcloud + Onyx-indexed metadata) by default.

IDP pipeline reference

NextcloudIntake & sync

WebDAV inbox receives uploads; processed files sync back to team folders.

Docling / TikaParse & extract

Layout-aware OCR for forms and W-2s; Tika handles 1,000+ text formats.

GotenbergNormalize to PDF

Office and HTML sources become consistent PDFs for downstream redaction.

StirlingRedact & fix

PII redaction, split, and merge before documents enter the archive.

ClawQL Archive LayerManaged archive

Nextcloud + Postgres metadata store, Onyx-indexed — the default on managed accounts. Paperless-ngx compatible for self-hosted deployments.

OnyxEnterprise search

Hybrid search and ingestion API — agents query indexed content via knowledge_search_onyx.

ConeshareSecure sharing

VDR share links and viewer webhooks for external parties who need controlled access.

Security

Built to prove, not just claim.

ClawQL documents how container images are scanned, signed, and enforced from CI through Kubernetes admission — plus a 32-module curriculum for agentic AI deployments. Same controls on self-hosted Helm and dedicated managed accounts.

Read the security hub

Golden image pipeline

OSV-Scanner, Trivy, and Syft SBOM gates run before any image publishes. One BuildKit build per image — the exact OCI layout scanned is what Cosign signs and pushes to GHCR. Failed scans block push, sign, and tag promotion.

Pipeline overview

Admission enforcement

The Helm chart defaults Kyverno verifyImages for clawql-mcp and clawql-website. Unsigned or unverified digests do not schedule — deployed ClawQL images tie back to the signed artifacts from CI, not ad-hoc kubectl.

Defense in depth

Built for agentic AI

32-module security curriculum from supply chain through runtime. Agentic Gateway ATR scoping limits what agents can do regardless of prompt injection; audit trails, sandbox isolation, and distroless images are documented for self-hosted k3s.

32-module curriculum

Prove it yourself

Security documentation is public: golden-image walkthroughs, deliverables matrix, npm supply-chain hardening, and cosign verify instructions. Operators and reviewers can reproduce CI gates — we document limits as clearly as controls.

Full security hub

Build → registry → cluster

Layer	What happens
GitHub Actions	Failed OSV/Trivy/SBOM or image scan → no push, no sign
GHCR + Sigstore	Cosign keyless signatures bind to the digest that passed CI
Kubernetes	Kyverno verifyImages rejects unsigned ClawQL images at admission

Questions & Answers

Start your 14-day trial or self-host free

No perpetual free hosted plan — self-host free forever on Apache 2.0, or start a 14-day Developer trial. Gateway tiers run on multi-tenant Regional Hubs with unlimited MCP executions, vault memory with no egress penalties on recall, and Onyx search on Teams. IDP tiers (Starter $299+) activate document processing, VDR, and sovereign inference on a dedicated tenant. One Agentic Gateway endpoint for every tier: upgrade from Teams to Starter without changing your URL, auth token, or vault history.

Self-hosted

$0

Want free forever? Run the full Apache 2.0 stack on your hardware — Helm chart, GHCR images, no license fee, no feature restrictions. Enable only the plugins you need via CLAWQL_ENABLE_* flags.

search, execute, audit, cache (Core — always on)
memory_ingest & memory_recall (default on)
Full IDP pipeline when you opt in (8 vendors)
Apache 2.0 — free forever, you pay infra only

Quick start

Gateway + memory · 14-day trial

Developer

$29/mo

Agentic Gateway + agent memory vault for developers connecting Claude Code, Cursor, or Codex to your APIs. Start with a 14-day free trial — no credit card. No IDP, no GPU inference.

Unlimited MCP executions
Regional Hub–hosted Agentic Gateway endpoint
Vault storage — no egress penalties on memory recall
One MCP endpoint on every tier — same URL when you upgrade

Start free trial

Gateway + memory + search

Teams

$99/mo

Full Onyx semantic search + memory vault + Agentic Gateway for teams building agent workflows. Still no IDP bundle — add Starter when you need document processing.

Unlimited MCP executions
Regional Hub–hosted Agentic Gateway endpoint
Vault storage — no egress penalties on memory recall
One MCP endpoint on every tier — same URL when you upgrade

Join early access

IDP plugin bundle

Starter

$299/mo

Activates the IDP plugin bundle on dedicated tenant infrastructure: Tika, Gotenberg, Stirling, archive layer, classify/extract, Coneshare VDR, sovereign inference. Your MCP endpoint and vault memory stay the same.

Unlimited MCP executions
Dedicated tenant · full IDP pipeline
5,000 documents/month
5 users · 50 GB storage

All tiers & limits

Need enterprise SLAs or on-call support?

Large enterprises and regulated industries. Dedicated node, custom fine-tune with retraining, multi-region (EU available), DPA/BAA, dedicated CSM. Sovereign Security Pack included.

Contact sales