Commercial Agent Tools: November 4, 2026

1. Droneship Runs Claude Code or Codex in gVisor Cloud Sandboxes Behind a Gateway That Keeps the Credentials

Droneship, posted to Show HN on November 4, starts Claude Code or Codex in an isolated cloud machine with droneship run claude or droneship run codex (Copilot, Gemini and Cursor are listed as coming soon), cloning the current repository and branch into a gVisor sandbox that boots in under a minute with Python, Node, Go, git, gh, npm, pnpm, pip and uv preinstalled. The agent holds no connected credentials: each run gets a SPIFFE identity, and an Agent Gateway matches network access on host names (deny by default, with one-click allows in the console), allows only the MCP tools turned on for the project, injects keys into allowed requests and checks model calls with Highflame’s guardrails. Every file touched, command run and tool call is logged for droneship logs, each repository keeps a saved workspace (uncommitted work and Claude’s past conversations included) that droneship resume claude reopens, and droneship diff and droneship pull bring changes back to a local branch. Claude Code routines can run on Droneship machines through a Claude self-hosted environment, which needs a Claude Team or Enterprise organization. Pricing: free during early access with up to 5 concurrent sandboxes and 30 sandbox hours a month using the user’s own Claude or ChatGPT plan or API key; a Pro tier with unlimited sandboxes, larger machines and always-on sandboxes is sold through Highflame on request. Source Show HN

2. Deputy Turns the Forms on Any Web Page Into Typed MCP Tools That Run in the User’s Logged-In Chrome

Deputy from Razin, posted to Show HN on November 3, is a Chrome extension plus npx deputy-mcp server that gives Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI, OpenCode and other MCP clients a typed API for the current page instead of screenshots or accessibility trees, and performs the call in the user’s real logged-in session. It applies Chromium’s WebMCP standard on sites that have not adopted it by writing toolname and description attributes onto the page’s own forms so the browser generates the JSON Schema and runs the submission, exposes tools that CopilotKit apps already register, and builds tools from ARIA widgets where no form exists; forms that change something wait for the user’s approval if the client can ask. In Razin’s own measurement on an admin page with 400 orders (Claude Sonnet through claude -p, median of three runs), Deputy needed 3 agent turns, 72,632 tokens and $0.072 against 10 turns, 245,250 tokens and $0.189 with Playwright MCP. It uses synthetic events rather than chrome.debugger, so heavy editors such as Google Docs may ignore its input. Pricing: Free with 10 tools and no hosted models; Pro at $9.99 a month or $95.99 a year adds 17 more tools and hosted models for browser_look, browser_ask and delegate_goal (5,000 a month, 500 a day). A source-available version with the 10 free tools is on GitHub under the PolyForm Shield 1.0.0 license, which is not an open-source license. Source Show HN

3. Chainguard Publishes 59,419 Version-Exact Manual Pages With a Five-Tool MCP Server for Agents

Chainguard launched man.cgr.dev on November 3, publishing every manual page in Chainguard OS (59,419 documents from 1,286 packages across its three package repositories), rebuilt nightly from the signed apks that make up the distribution, so agents working in minimal container images without -doc packages can look up flags for the exact version they run. Each page has a pinned address per package version (or latest) and comes as HTML, Markdown (.md), original roff (.raw) or a JSON manifest with content digests (.json), and 91% of cross-references resolve to a real page. An MCP server at man.cgr.dev/mcp exposes search_docs, get_man_page, resolve_xref, list_package_docs and get_package_info. Pages enter only from signed apks, pages under BUSL, SSPL, Elastic or proprietary licenses are withheld, and the site deliberately admits AI crawlers through permissive robots.txt, sitemaps and llms.txt. Pricing: free, with every page public and no Chainguard account required, including pages for enterprise packages. Source Show HN

4. Conecto Gives Agents 581 Hosted Business Tools Behind One MCP Server, Paid Per Call From One Balance

Conecto, posted to Show HN on November 3, is a hosted set of 581 tools across 52 business areas (among them 41 CRM tools, 33 bookkeeping, 24 invoicing, 13 calendar, 10 email and 8 phone tools, plus help desk, e-signature, HR, forms and newsletters) that an agent such as ChatGPT, Claude, Grok Bot or Hermes Agent reaches through one MCP server, a REST API with OpenAPI, or a SKILL.md file it reads to sign itself up in one call. Conecto hosts the servers, database, phone lines, mail and file storage, and can also work inside an existing HubSpot, Salesforce, Microsoft 365 or Google Workspace account; every call is logged with the key that made it and its cost, owners can set rules so refunds or deletions wait for approval, and company sign-in and Okta or Entra ID integration are supported. Pricing: usage-based from a single prepaid balance, with record changes at $0.01, reads free up to 1,000,000 a month and an inbox at $1 a month; agents start with $1 of credit, top-ups start at $5 and the first is doubled up to $20. Conecto’s own comparison puts one person’s month of CRM, help desk, scheduling, e-signature and accounting work at $25.10 against $98 for five SaaS subscriptions. Source Show HN

5. Cinnamon Is a Household Agent With Shared and Private Spaces That Asks Before Every Send, Booking or Purchase

Cinnamon, posted to Show HN on November 3, is a consumer agent that a person can use alone or share with the people they invite (the site’s main pitch says up to five, and the larger plans allow up to 10 members), with a household space everyone sees (trips, chore rotas, a Sunday digest) and a private space per member that it keeps separate. It plans trips, digests a connected Gmail inbox, tracks appointments, forms and renewals, watches prices against a target, and drafts emails into Gmail Drafts; every email, booking, purchase or share stops for approval scoped to one time, one task or 7 days, passwords and cards sit in a vault the agent cannot read (it sees stand-in tokens), each task runs in its own sandbox, and every action lands in a timeline that can be exported or deleted. The FAQ says it never sends email, edits the calendar or spends money on its own and does not keep conversation history after a session. Pricing: 25 free trial credits; Personal at $9.99 a month for 50 credits; Group at $19.99 a month for 150 shared credits and up to 6 members; Group Max at $99.99 a month for 1,000 credits and up to 10 members. Source Show HN

6. Hookdeck Inspect Is a No-Account Webhook Tester an Agent Can Drive With Two HTTP Calls

Hookdeck posted Inspect to Show HN on November 3, a webhook tester built for coding agents as well as people: POST https://inspect.hookdeck.com/ returns an ingest_url to give the webhook provider and a requests_url that long-polls for up to 25 seconds and returns the next request as JSON, so an agent can register a webhook, trigger it and read the verbatim payload without a browser. It needs no account, keeps requests on the free tier, publishes an OpenAPI description, ships as an agent skill (npx skills add hookdeck/agent-skills --skill inspect) and over MCP, and offers mock payloads from more than 450 platforms; the page tells agents that received payloads are data from others, not instructions. Requests can be forwarded to localhost with the Hookdeck CLI. Pricing: free; Hookdeck’s paid Event Gateway (with a free starting tier) adds queues, retries and delivery logs for production. Source Show HN

7. PagerDuty’s Paige SRE Agent Can Go On Call, With Per-Team Autonomy Limits and a Shared Incident Memory

PagerDuty introduced Paige, its next-generation site reliability engineering agent, and renamed PagerDuty Operations Cloud to the PD Reliability Platform. Teams set Paige’s permissions individually: with a person in the loop it triages and holds remediation for approval, and it acts alone only in low-risk scenarios a team has explicitly authorized; it can join escalation policies as an on-call responder, so some lower-priority incidents never reach a human. Recommendations arrive with evidence and a confidence level, resolved incidents are written to a shared memory Paige consults on similar failures, post-incident reviews are generated automatically with owners and due dates for follow-ups, and a new console shows which agents are active and what each is doing. PD Labs opens three previews to design partners: Topology (a live service map built from code, infrastructure and runbooks, with no CMDB), Deep Investigation (parallel root-cause hypotheses with visible reasoning) and SRE Workbench (a shared human and agent workspace). PagerDuty says a benchmark report released with the launch shows Paige beating leading AI coding agents at diagnosing simulated IT incidents; pricing was not disclosed. Source

8. Snowflake Adds a Business Ontology for Its CoCo and CoWork Agents and Expands Cortex AI Gateway Governance

Snowflake announced agent features at its Expedition 2026 virtual conference. Business Ontology, an extension of its semantic views in private preview with public preview coming, lets organizations set approved business definitions that its data-team coding agent CoCo and its knowledge-worker agent CoWork use to ground answers, so two agents do not give incompatible answers to the same question. Cortex AI Gateway, introduced in July as a central point to govern and monitor agents, gains agent abuse detection, observability and tracing, Cortex sandboxes, AI cost governance, dynamic model routing and tool integration built on its Natoma acquisition, applying existing data masking and access restrictions to agent actions and letting IT restrict which external tools agents can reach. Snowflake also announced autonomous data engineering capabilities that let agents build, monitor and troubleshoot data pipelines; pricing was not stated. Source