ClawQL

Local-first MCP server for token-efficient AI agents. Search/Execute against OpenAPI w/graphql. Knowledge graph memory + document OCR. Make your data queryable.

[ClawQL @ClawQL]

Istio local: MCP routes fixed under clawql.localhost; 127.0.0.1 MCP URL guidance
Egress allowlist + runtime class scaffolding (Kata / gVisor + Kyverno)

Gitleaks + scheduled TruffleHog; npm ci reliability on Node 25

Links: npm clawql-mcp · https://t.co/HSo9C9JF5M

(End of 🧵) 🦞

[ClawQL @ClawQL]

Shipped clawql-mcp 6.1.0

This release is mostly about furthering our defense-in-depth security posture. All secrets now injected from vault (with a locally hosted dashboard for easy updates), image signing enforced, and a new MCP proxy sitting in front to intercept attacks.

🧵

[ClawQL @ClawQL]

Highlights:

Helm: Vault subchart + secretSourcing posture; External Secrets + Vault KV docs

Dashboard: optional bundled Vault-first UI + Kyverno image verify; docs.localhost vs clawql.localhost

Panguard bridge: mcpProxy, optional JWT, unary gRPC MCP delegation

🧵(cont)

[Daniel Smith | Building ClawQL Agents @DanielSmithDev]

Full announcement of clawql-ouroboros below ⬇️🫡

Excited to share a new open-source release:

clawql-ouroboros — a lightweight TypeScript library for specification-first evolutionary loops in agent and automation backends.

https://t.co/QpoRh0q10q

The Problem It Solves

Most LLM-powered agents or automation workflows go beyond a single prompt. You often need:

• A frozen specification (goals, acceptance criteria, structured ontology/schema)
• Real execution against APIs, systems, or tools
• Evaluation of whether the output actually meets the bar
• Evolution of the spec/plan when it falls short — then loop until convergence

Without structure, these loops become brittle, token-heavy, and hard to reason about.

How clawql-ouroboros Works

It provides the reusable skeleton for a “run → evaluate → evolve” pattern:

1. Seed — Defined with Zod schemas: immutable goal, acceptance criteria, ontology (structured output fields), constraints, etc.
2. EvolutionaryLoop — The core orchestrator.
3. Wonder/Reflect — Hooks (often LLM-powered) to generate insights and suggest refinements.
4. Executor — Your custom side effects (REST, SDKs, sandboxes, etc.).
5. Evaluator — Scores results against criteria (rules, tests, LLM-as-judge, or hybrid).
6. ConvergenceCriteria — Built-in gates for similarity, stagnation, oscillation, regression, plus min/max generation limits.
7. EventStore — Tracks lineage (InMemoryEventStore for quick starts; implement your own for Postgres/Redis).

You wire in your own LLM adapters and real executors — the library stays focused and embeddable. No bundled database or external LLM calls by default.

Minimal conceptual flow (simplified):
• Start with a Seed
• Execute → Evaluate
• If not converged → Reflect → Evolve Seed → Repeat
• Stop on convergence or caps

Full examples, including MCP hooks and a background poller, are in the docs.

Relation to Q00/Ouroboros

This is not a full port of the excellent Python Q00/ouroboros project (which includes CLI, interviews, persistence, etc.). It’s a scoped, embeddable TypeScript subset focused on the typed evolutionary core — perfect for dropping into Node services.

Tech Details
• Node.js ≥ 22
• Dependencies: Zod ^4 + uuid ^11
• License: Apache-2.0
• Repo: https://t.co/TsjuB15Sk2
• Docs: https://t.co/57PSfaJqbo

Who It’s For

• Agent/platform engineers building consistent, typed LLM loops
• Internal tool teams with existing execution layers
• Anyone maintaining TS services at the LLM + API + verification boundary

If you’re tired of reinventing generation/evaluation/convergence logic per project, this might save you time.

Feedback, issues, or contributions welcome on GitHub!

[Daniel Smith | Building ClawQL Agents @DanielSmithDev]

We just shipped clawql-mcp 4.1.0 — a meaningful step for anyone running Model Context Protocol against real OpenAPI / Google Discovery surface area.

ClawQL is not “yet another API wrapper.” It is a spec-first MCP process where optional capabilities stay opt-in, the GraphQL layer keeps answers lean, and the vault and tools you already run (Obsidian, Slack, document APIs) stay composable. Version 4.1.0 is a step toward the unified narrative in our public deck: APIs, documents, memory, and knowledge in one operable surface — without giving up self-hosting or clear boundaries between features.

What is in the box (all opt-in, same binary):

    •    knowledge_search_onyx — When you enable Onyx in your merge and set CLAWQL_ENABLE_ONYX, ClawQL exposes a thin knowledge_search_onyx tool that routes to your self-hosted Onyx (https://t.co/nVFUa7G4OT) send-search-message flow. The same execute stack applies; you keep a single spec-driven mental model. The bundle also includes an Onyx ingestion path for execute workflows (post-archive indexing), documented in the repo. This is the “enterprise knowledge next to your APIs” chapter we have been building toward.

    •    notify — Turn on CLAWQL_ENABLE_NOTIFY and you get a first-class Slack chat.postMessage wrapper, aligned with the bundled Slack spec and the same error surfaces as the rest of execute. It is the obvious “done” and “failed” channel for long-running or multi-step automation — without shelling out to a separate process.

    •    Ouroboros — The new clawql-ouroboros workspace package (also on npm at 0.1.0) is an evolutionary loop library. When you set CLAWQL_ENABLE_OUROBOROS, the server registers three ouroboros_* tools backed by the library, with an optional Postgres-backed event store for lineage when you configure

    •    Memory — memory_ingest can carry optional enterprise citation payloads into the vault, and the existing toolOutputsFile path remains the escape hatch for very large bodies without stuffing megabytes into MCP tool JSON (added in the previous line release, now part of the story customers tell when ingesting big decks and logs).

Why we care: 
Feature flags keep the default surface small; stdio, Streamable HTTP, and gRPC stay the same wire story. GraphQL in-process still trims execute responses. Helm 0.4.0 / app 4.1.0 for operators.

If you are wiring AI assistants to real systems, the hard part is rarely “call an API once.” It is keeping context small, auth sane, and workflows coherent across sessions. ClawQL 4.1.0 pushes that story forward: optional tools that sit beside the two core tools — search and execute — without turning the MCP into a monolithic “everything in one JSON blob” server.

Links: 
GitHub release: https://t.co/Xb0JyufXe4
npm: [email protected]
Docs: https://t.co/4O23s1cMXX

If you are building agent platforms on top of MCP, this is a good release to pin.

#MCP #OpenAPI #ClawQL #AIinfrastructure #opensource

[ClawQL @ClawQL]

Full announcement:

https://t.co/q9lWkPGqWs

[Daniel Smith | Building ClawQL Agents @DanielSmithDev]

We just shipped clawql-mcp 4.1.0 — a meaningful step for anyone running Model Context Protocol against real OpenAPI / Google Discovery surface area.

ClawQL is not “yet another API wrapper.” It is a spec-first MCP process where optional capabilities stay opt-in, the GraphQL layer keeps answers lean, and the vault and tools you already run (Obsidian, Slack, document APIs) stay composable. Version 4.1.0 is a step toward the unified narrative in our public deck: APIs, documents, memory, and knowledge in one operable surface — without giving up self-hosting or clear boundaries between features.

What is in the box (all opt-in, same binary):

    •    knowledge_search_onyx — When you enable Onyx in your merge and set CLAWQL_ENABLE_ONYX, ClawQL exposes a thin knowledge_search_onyx tool that routes to your self-hosted Onyx (https://t.co/nVFUa7G4OT) send-search-message flow. The same execute stack applies; you keep a single spec-driven mental model. The bundle also includes an Onyx ingestion path for execute workflows (post-archive indexing), documented in the repo. This is the “enterprise knowledge next to your APIs” chapter we have been building toward.

    •    notify — Turn on CLAWQL_ENABLE_NOTIFY and you get a first-class Slack chat.postMessage wrapper, aligned with the bundled Slack spec and the same error surfaces as the rest of execute. It is the obvious “done” and “failed” channel for long-running or multi-step automation — without shelling out to a separate process.

    •    Ouroboros — The new clawql-ouroboros workspace package (also on npm at 0.1.0) is an evolutionary loop library. When you set CLAWQL_ENABLE_OUROBOROS, the server registers three ouroboros_* tools backed by the library, with an optional Postgres-backed event store for lineage when you configure

    •    Memory — memory_ingest can carry optional enterprise citation payloads into the vault, and the existing toolOutputsFile path remains the escape hatch for very large bodies without stuffing megabytes into MCP tool JSON (added in the previous line release, now part of the story customers tell when ingesting big decks and logs).

Why we care: 
Feature flags keep the default surface small; stdio, Streamable HTTP, and gRPC stay the same wire story. GraphQL in-process still trims execute responses. Helm 0.4.0 / app 4.1.0 for operators.

If you are wiring AI assistants to real systems, the hard part is rarely “call an API once.” It is keeping context small, auth sane, and workflows coherent across sessions. ClawQL 4.1.0 pushes that story forward: optional tools that sit beside the two core tools — search and execute — without turning the MCP into a monolithic “everything in one JSON blob” server.

Links: 
GitHub release: https://t.co/Xb0JyufXe4
npm: [email protected]
Docs: https://t.co/4O23s1cMXX

If you are building agent platforms on top of MCP, this is a good release to pin.

#MCP #OpenAPI #ClawQL #AIinfrastructure #opensource

[ClawQL @ClawQL]

ClawQL 4.1.0 is out. Same MCP: search + execute on OpenAPI / Discovery. New optional tools sit behind flags — you choose what loads.

🧵1/12

[ClawQL @ClawQL]

Thread links: CHANGELOG (https://t.co/disq4A4AqH) · mcp-tools (https://t.co/TZmu0Om1kv) · https://t.co/HSo9C9JF5M (https://t.co/fz6LmhLqPo) — DMs open for production feedback.

🧵12/12