Open Source Agent Tools: October 22, 2026

1. Omnara Pitches a Self-Hostable Alternative to Claude Managed Agents and OpenAI’s Agents API

Omnara (Omnara). Omnara presented its open-source managed agent platform on Show HN: an API that runs agents defined in a YAML config on any model (OpenAI, Anthropic, OpenRouter or self-hosted) and any machine, with tools, MCP, skills, secrets, approvals and event streaming built in. Machines can be your own laptop, VM or container, or sandboxes from providers such as Daytona, E2B, Modal and Unikraft, and each agent’s event log lives in Postgres. It ships a CLI and a TypeScript SDK, and self-hosting is a single Docker Compose command; the hosted Omnara Cloud is free with your own keys and machines or pay-as-you-go for its models and sandboxes. The project shipped daemon and cluster v0.2.9 releases on October 21. License: Apache 2.0. GitHub

2. Union.ai’s Hlock Pins an Agent Harness Into a Lockfile You Can Commit

hlock (Union.ai). hlock applies the package-manager lockfile model to agent setups: hlock pin records the harness (Claude Code, Codex, Pi, OpenCode, Cursor, Gemini CLI and others), models, MCP servers, skills and config into a hashed harness.lock, with secrets stored as placeholders. hlock sync rehydrates the setup on another machine or in Docker, hlock launch starts an isolated session on the locked setup, and hlock audit checks integrity, hidden characters and secrets. Teams can commit per-person locks to try a colleague’s setup, and an eval-pack command scaffolds project-specific tasks for comparing stacks. The Rust core ships as wheels on PyPI and a crate (v0.5.0 released October 21; the project calls itself early development). License: Apache 2.0. GitHub

3. Twine Is a Native macOS Workspace for Multi-Agent Coding Workflows

Twine (Standing Desks). Twine wraps existing agent CLIs (Codex, Claude Code, Pi, OMP, Antigravity and OpenCode v2) in a macOS app that runs them as workflows: a single agent, an adversarial implementer-and-reviewer pair, a coordinator that splits a plan across two workers, or a custom agent graph built in a visual editor and saved as TOML. A Traces panel records prompts, tool calls, model and usage details per step, and memories are shared across agents. It requires macOS 26 on Apple Silicon and installs via Homebrew; usage is billed through each harness’s own subscription. License: MIT. GitHub

4. Stalemap Tells Coding Agents Which Dependencies Changed After Their Training Cutoff

Stalemap. Stalemap compares the dependency versions a model likely knows with the installed versions, diffing type declarations, release notes and upgrade guides, and writes a short ranked list of removed or renamed APIs and changed defaults (each with a replacement) into CLAUDE.md or AGENTS.md. A guard mode runs the project’s tsc after each edit and annotates errors caused by changed dependencies, and a search_changes MCP tool lets agents query on demand. It covers JavaScript and TypeScript projects only and scans code locally, sending only package names and versions to npm and GitHub. License: MIT. GitHub

5. Flint Gives Agents Reversible Context Compression and Version-Bound Edits

Flint. Flint is a standard-library Python tool (CLI and JSON interface, no model or server) that compresses code, JSON and text into reversible encodings, packs selected sources under a byte budget and stores searchable decisions that expire when their linked source files change. Edits are bound to the file revision the agent read, so stale patches are rejected, and incremental rereads return only changes since an acknowledged version. Its demo cuts 12 minified JSON records from 1,226 to 499 bytes with full recovery, which the author notes is a byte count rather than a token-savings guarantee. It targets Python 3.11+ on Linux or WSL. License: Apache 2.0. GitHub

6. Renor Shares Coding-Agent Memory as Local Markdown Files Over MCP

Renor. Renor stores what you tell Claude Code, Codex or Cursor as Markdown notes in a local folder and serves them back to any of those agents through an MCP server, so a decision saved in one agent’s session can be recalled in another’s. It runs locally on Node 22 with no account or API key and sends nothing off the machine unless a model is configured; a hosted version is available at renor.cloud. License: AGPL-3.0. GitHub

7. QAmcp Tests MCP Servers for Hidden Instructions, Leaked Secrets and Missing Safety Labels

QAmcp. QAmcp is a CLI and GitHub Action that connects to an MCP server (remote or stdio) the way an AI client would and reports deterministic pass/fail results: hidden instructions in tool descriptions, secrets leaked in error output, slow or hanging tools, unhandled bad input, missing safety labels and OAuth setup. Tools that modify data are never run, and read-only tools without labels run only with an explicit --owner flag. No LLM judging is involved and nothing is sent anywhere except the server under test. License: MIT. GitHub

8. Agent Middleware Puts Signed Permits and Receipts in Front of Risky Agent Tool Calls

Agent Middleware. Agent Middleware is a gateway for consequential MCP tools such as refunds, payouts and deploys: each call needs a signed permit scoped to specific tools, a budget and an expiry, runs at most once per idempotency key, and produces a signed receipt with request and response hashes that can be verified offline. The README is explicit about limits: receipts cannot prove a remote effect happened, and uncertain deliveries must be reconciled manually rather than retried. It is pre-release, solo-maintained and installed from source. License: Apache 2.0. GitHub