Open Source Agent Tools: October 30, 2026

1. Wavefunction Is a Minimal Desktop Coding Agent for macOS That Runs Computer Use in Its Own Container

Wavefunction. Wavefunction, posted to Show HN on October 29 with v0.2.0 shipped the same day, is a desktop coding agent for the Mac that deliberately avoids the “superapp” direction of other agent clients. It splits tasks across subagents that the user can steer mid-task, giving any subagents that could touch the same files their own git worktree, and lets the user fork a chat from any message or rewind both code and conversation. Computer use runs inside a local Docker container with its own desktop rather than on the user’s screen, and it ships 70+ MCP connectors, reads AGENTS.md and SKILL.md files, and reports token, cache and provider cost usage. Models come from a ChatGPT sign-in, API keys for Anthropic, Gemini, OpenAI, OpenRouter or Vercel AI Gateway, or local models via Ollama and OpenAI-compatible servers, with an option to require zero-data-retention models. v0.2.0 adds provider-native web search for Anthropic, Gemini and Vercel AI Gateway models. It is built on LangGraph and packaged with Tauri. License: MIT. Source

2. Cornercase Turns an Issue Into a Coding Agent Working in Its Own Worktree and Tab

Cornercase. Cornercase, posted to Show HN on October 29 alongside three releases that day (up to v0.12.227), is a mouse-driven terminal multiplexer for running coding agents in parallel. Opening an issue from GitHub, GitLab, Linear, Jira, Plane, Asana or Shortcut and pressing start creates a branch and worktree named after the issue, opens a tab and hands the issue to the agent as its first prompt; it supports Claude Code, Codex, Gemini, Grok, Cursor, Copilot and more than a dozen others. Tabs show which agents are working, waiting or done along with the model and context fill, a changes panel shows the live diff next to the agent with per-hunk staging, discard or send-back, and a background server keeps shells alive after the window closes so sessions can be reattached from another terminal or over SSH. A cornercase CLI lets one agent drive the others. It is built in Rust on Ghostty’s terminal core and installs via Homebrew or a shell script on Linux and macOS. License: MIT OR Apache-2.0. Source

3. phpClaw Brings a ReAct Agent Runtime With Prompt-Injection Guards to PHP Frameworks

phpClaw. phpClaw, posted to Show HN on October 29, is an agent engine for PHP 8.1+ whose core has no framework dependencies beyond ext-curl, ext-json and ext-mbstring. A single builder picks up whichever provider key is present (Anthropic, OpenAI, Groq, Gemini, Mistral, DeepSeek or an Ollama host) and runs a ReAct loop with a hard iteration cap over pluggable tools such as an allowlisted shell, HTTP, file reads, database queries and code search. It adds eight default prompt-injection guards, swappable memory drivers, on-demand skills, Pipeline and Graph primitives for chains and multi-agent graphs with human pause and resume, and a bundled MCP server that exposes a PHP app to MCP clients. Adapters cover Laravel, Symfony, WordPress, Drupal, Magento, Joomla, OpenCart and PrestaShop. It deliberately excludes RAG, embeddings and vector stores. Install with composer require phpclaw/phpclaw. License: MIT. Source

4. Apgar Flags When a New Agent Release Behaves Worse Than the Last One, Using Only Traces

Apgar. Apgar, from oimo-ai, watches the first hours after each agent deploy and compares it with the previous release’s baseline on signals such as tool error rate, retries and loops, tool call counts, answer length, cost and latency, then sends one Slack message or webhook per release naming the signal that moved and where the change is concentrated. It needs no LLM judge, no dataset and stores no prompt or completion text, working from OpenTelemetry traces or traces pulled from Langfuse, with each deploy tagged by a release attribute. It self-hosts on Docker Compose with ClickHouse, and v0.1.3, released and posted to Show HN on October 29, adds a Helm chart for Kubernetes plus an agent skill (npx skills add https://docs.useapgar.dev) for having a coding agent wire up the instrumentation. Note that v0.1.3 also added telemetry sent to PostHog. License: Apache-2.0. Source

5. Charter 0.6 Enforces Per-Field Rules on Agent Calls Into Customer Accounts and Adds OAuth Scope Handling

Charter. Charter, from r28ai, is a Python library for agents that act in customers’ accounts through APIs: each tool is a Pydantic model mirroring the API field by field, every call is validated against it before it is sent, and fields marked FromContext() (such as the customer ID) are filled by the application per call so the model cannot change whose account it acts on. The same declarations generate an egress map listing what the model sees and what is withheld, which can be committed so access changes show up as diffs in code review. v0.6.0, released and posted to Show HN on October 29, adds AnyOf scope declarations and missing_scopes for checking what an OAuth grant still lacks, Tool.derived(scopes=...) for narrowing a tool’s consent, and a unified InsufficientScope error that normalizes scope refusals from Google, Slack, Linear, GitHub, Shopify and others. It runs in-process with no proxy or telemetry and depends only on pydantic and httpx. Install with pip install charter-ai. License: Apache-2.0. Source

6. Mnestiq Records Signed, Tamper-Evident Evidence of What an Agent Read, Decided and Did

Mnestiq. Mnestiq, posted to Show HN on October 29 after its v1.0.0 release on October 28, is a “flight recorder” for AI agents that tags every piece of model context with its provenance (system, user, tool output, retrieved document, web, agent message, memory) so prompt-injection incidents can be traced to the step and untrusted input that preceded them. Records are hash-chained and sealed with Ed25519-signed Merkle checkpoints so edits, deletions or reordering fail verification, prompts and tool data can be redacted for auditors while the chain still verifies, and outbound HTTP and raw TCP connections are logged with IPs and ports to join against firewall logs. It records either through a no-code proxy in front of the Anthropic or OpenAI APIs (also Azure OpenAI, Ollama, vLLM and LiteLLM) or by wrapping the Python SDK clients, and ships a local dashboard. Version 1.0 stabilizes the evidence format, adds per-value salted digests, delta-only recording of model context and Microsoft Entra Agent ID tagging. Install with pip install mnestiq. License: Apache-2.0. Source