Self-Hosting Day: OpsCrew Builds the Server Side While Guardrails Redact and Cap Spend

Today's findings

1. #1 OpsCrew deploys repos to your own server from Claude Codetool

   An open-source CLI and Claude Code plugin takes a GitHub repo through Docker, TLS, DNS, CI/CD, and monitoring setup onto your own infrastructure.

   Open-source DevOps tool

   Repo in, self-hosted app out

   1

   Repo

   GitHub project

   2

   Docker

   Containerize

   3

   TLS

   HTTPS certs

   4

   DNS

   Domain routing

   5

   CI/CD

   Build + deploy

   The apply step — review the generated config before it runs

   6

   Monitor

   Uptime checks

   OpsCrew, a CLI + Claude Code plugin, generates each stage for your own infrastructure.

   Why it matters: AI-assisted coding makes app creation faster, but self-hosted deployment remains the manual bottleneck for small teams.

   How to apply: Add OpsCrew to a staging repo, review each infrastructure step before apply, then codify the generated config into your own runbooks.

   devopsclaudedeploymentopen-source

   Read more: I used Claude Code to build an open-source DevOps tool that deploys your apps to your own server

2. #2 CleanWeb MCP trims web-scrape noise and injection risktool

   An MCP server for Claude Desktop and Cursor strips boilerplate from scraped pages, cuts token load, and blocks common prompt-injection patterns.

   Agents browsing untrusted pages

   Straight into context, or screened at the gate

   Direct fetch

   vs

   Through CleanWeb MCP

   Route agent fetches through the MCP server, then regression-test against known injection payloads.

   One server between the scraped page and the agent's context.

   Why it matters: Agents that browse untrusted pages otherwise ingest noisy content and hidden instructions directly into context.

   How to apply: Attach the MCP server, route web fetches through it, compare token counts, and add regression tests with known injection payloads.

   mcpsecurityscrapingtokens

   Read more: [Showcase] I built a CleanWeb MCP server that cuts web scraping token noise by 87% and blocks prompt injections for Claude Desktop & Cursor

3. #3 Cloakwall redacts PII and secrets from LLM traffictool

   A stdlib-only LiteLLM guardrail redacts sensitive entities from prompts, responses, streams, and tool calls while writing a tamper-evident audit log.

   LLM privacy guardrail

   LLM calls pass — but only redacted

   bounded capability

   LLM traffic in flight

   scope Prompts & responses

   limit Streamed chunks & tool calls

   monitor Tamper-evident audit log

   Stdlib-only LiteLLM guardrail or proxy — keep the audit anchor off the log host.

   Why it matters: Teams shipping LLM features need a request-path control for privacy and secret leakage, not just prompt instructions.

   How to apply: Deploy as a LiteLLM guardrail or proxy, test streamed chunks and Anthropic tool blocks, and store the audit anchor outside the log host.

   securityprivacyguardrailslitellm

   Read more: Cloakwall: stdlib-only PII redaction and tamper-evident audit logging for LLM traffic(Guardrail) · Looking for testers and Feedback:Cloakwall, a PII/secret-redaction proxy and LiteLLM guardrail for LLM traffic

4. #4 Circuit breaker caps cumulative agent spend and loopstool

   A drop-in breaker tracks retries and tool-call loops against budgets that survive new root runs, instead of per-run limits that reset.

   Agent guardrails

   Per-run limits reset — the breaker keeps counting

   vs

   Per-run limits

   Circuit breaker

   Counter reset

   Wipes each new run

   Survives new root runs

   Retry spend

   Looks small per run

   Accumulates toward budget

   Tool-call loops

   Reworded, clock restarts

   Counted per workflow

   Threshold trip

   Never trips

   Fail closed + alert

   Per-run limits wins the row Circuit breaker wins the row

   Runaway agents fail slowly — many small calls, one cumulative bill.

   Why it matters: Runaway agents often fail slowly through repeated retries and reworded queries, not a single giant request.

   How to apply: Persist spend and loop counters by workflow, wrap LLM and tool calls with the breaker, and fail closed with alerting when thresholds trip.

   agentscost-controlreliabilityguardrails

   Read more: Built a drop-in Circuit Breaker to prevent runaway LLM costs and infinite agent loops (Free code + calculator)

5. #5 Configure permissions for headless Claude Code scriptstip

   Recent Claude Code versions start claude -p and Agent SDK runs without a permission mode in Manual, so needed tool calls are denied unless you add allow rules, hooks, or a permission-prompt tool.

   Why it matters: Automation and CI jobs can fail at the first tool call even though the same command works interactively.

   How to apply: Add the narrowest allow rules for required commands, use hooks or canUseTool where needed, and test the script in a container before broad rollout.

   claude-codeautomationpermissionsci

   Read more: Your claude -p script says a tool call was denied? Three ways to fix it, and what each one lets Claude do

6. #6 Beads turns agent memory into a ticket databasetool

   Beads models agent handoffs and context as tickets with epics, dependencies, blockers, and states instead of replaying full transcripts.

   Why it matters: Structured tickets keep context clean and make multi-agent delegation auditable and resumable.

   How to apply: Create epics and ticket states for agent tasks, store decisions and open items in tickets, and have agents read only the relevant ticket subtree.

   agentsmemorycontextworkflow

   Read more: 90% of problems here are solved by beads.

7. #7 Claude Code subagents run in parallel git worktreestechnique

   PotionUI's build demonstrates splitting work across Claude Code subagents, each isolated in its own git worktree, then reviewing and merging.

   Why it matters: Parallel agents can accelerate independent tasks, but shared working directories create conflicts and muddy review.

   How to apply: Break work into independent branches, assign one worktree per subagent, enforce shared style and tests, and merge only after per-branch review and CI.

   claude-codeagentsgitworkflow

   Read more: PotionUI: an open-source generation studio built with Claude Code subagents working in parallel git worktrees

8. #8 Mini Benchmarks curates evals for local agentic codingrepo

   A curated benchmark selection targets local agentic coding so teams can compare models on tasks closer to real coding-agent work.

   Why it matters: Local model choices are often made from single prompts; repeatable evals reduce the risk of picking a model that fails in multi-step coding.

   How to apply: Run the suite against top local candidates, track pass rate, latency, and memory, and use results to define default and fallback models.

   evalslocal-llmcoding-agentsbenchmarks

   Read more: Mini Benchmarks - a curated selection of benchmarks for local agentic coding evals

9. #9 Video grep gives agents a terminal search for video momentstool

   An open-source CLI indexes video so you can describe a moment and get timestamps without a GUI.

   Why it matters: Agent harnesses that only have shell access can now inspect recordings and test videos as part of debugging or content workflows.

   How to apply: Index test footage, expose the search command as an agent tool, and return timestamped clips for review rather than raw video.

   agentsvideocliopen-source

   Read more: Video grep in terminal for Agents · Video grep in terminal for Agents

10. #10 Mage-VL-Cowork packages a 4B open VLM for computer usepaper

    Microsoft's 4B vision-language model uses workflow-first environment synthesis and teacher-harness evolution to handle hybrid computer-use tasks.

    Microsoft · Mage-VL-Cowork

    A 4B open VLM that drives the computer

    4B

    parameters — small enough to run locally

    open weights

    Open

    checkpoint on Hugging Face

    Hybrid UI tasks

    computer-use workflows

    Sandboxed

    keep agent actions contained

    Trained with workflow-first environment synthesis and teacher-harness evolution.

    Why it matters: Small open VLMs can lower the cost and privacy barrier for UI automation, especially when paired with sandboxed tool use.

    How to apply: Download the Hugging Face checkpoint, build a constrained test harness around common UI tasks, measure failure recovery, and keep actions sandboxed.

    papervlmcomputer-uselocal

    Read more: microsoft/Mage-VL-Cowork · Hugging Face

11. #11 DeepSeek-V4.1-Flash stack publishes multi-node local serving reciperepo

    A detailed vLLM stack with 2.75 bpw EXL3, tensor-parallel DGX Sparks, and an RDMA-offloaded 3090 box reaches 2,600+ tok/s prefill to 256K and 322 tok/s structured output.

    Why it matters: It gives teams a concrete reference for long-context local serving across mixed hardware, including quantization and offload details.

    How to apply: Review the published configs and raw data, reproduce on a smaller context first, validate RDMA and Engram assumptions, and benchmark structured-output workloads.

    local-llmvllminferencequantization

    Read more: Lawyer with too many GPUs, part 3: DeepSeek-V4.1-Flash on 2x DGX Spark + an old 3090 box. 2,600+ tok/s prefill to 256K, 322 tok/s structured output, full stack and raw data published · Lawyer with too many GPUs, part 3: DeepSeek-V4.1-Flash on 2x DGX Spark + an old 3090 box. 2,600+ tok/s prefill to 256K, 322 tok/s structured output, full stack and raw data published

12. #12 Clover-35B-A3B fits a 64 GB Mac with mixed-precision quantizationtip

    A mixed 6-bit checkpoint with selected BF16 tensors brings Clover-35B-A3B to about 29.4 GB and roughly 42 tok/s on a 64 GB Mac, though quality still needs work.

    Why it matters: It shows a reproducible way to fit a larger MoE model on Apple Silicon without full retraining, useful for local coding assistants and private inference.

    How to apply: Reproduce the mixed-precision recipe, retain sensitive tensors in BF16, benchmark long-output quality against your tasks, and treat current regressions as a gating issue before production.

    local-llmquantizationmacinference

    Read more: Clover-35B-A3B on a 64 GB Mac: 29.4 GB quantized weights and ~42 tok/s, with quality still unfinished