6 new indie AI agent tools from Aug 3–7: coding control, scheduling, guardrails, and local security

Six additional public builds fit this scan's August 3–7, 2026 window in Africa/Lagos. The previous issue covered HAR, Mirafold, Remembrane, Learning Mode, Modality, and CoArena; none of those projects appears here. Every pick has public source code and a concrete path to try it.

The common thread is control around the model. ADE and Wallfacer make coding-agent work easier to organize. DayOtter puts approval gates around calendar actions. Prime Agent keeps a long-running coding process alive. ModelFuzz checks tool calls before they run. Nightcrawler pushes a security agent onto a phone, where the safety boundary matters as much as the demo.

What shipped. ADE is an AGPL-3.0, local-first workspace for Claude Code, Codex, Cursor, Factory Droid, and OpenCode. Its repository describes desktop, web, terminal, and mobile clients, with a persistent "Brain" process that owns the project catalogue, synchronizes clients, and can run projects. Project data lives in .ade/; machine state lives in ~/.ade. 12

The useful distinction is between an agent and the place where its work is managed. ADE does not ask you to replace every coding CLI. It gives several of them a common project list and interface, which is a better fit for a builder who moves between a laptop, browser, terminal, and phone. The four-client approach is the novel part; the model remains the replaceable component. 2

Try it. The project site gives a Unix install command, curl -fsSL https://ade-app.dev/install.sh | sh, followed by ade code. The repository also links releases and Homebrew installation for the desktop app. 23

Caveat. The launch materials describe Linux desktop support as not yet available, Windows as beta and x64-only, and Android as forthcoming; iOS access is through TestFlight. That makes ADE most immediately useful to macOS users or people comfortable with its terminal and web paths. 1

What shipped. DayOtter is an open-source, self-hostable scheduling assistant that connects Google, Microsoft 365, Apple CalDAV, and ICS feeds. Its assistant, Otter, accepts natural-language requests but follows a confirm-first rule: it does not change a calendar until the user approves the action. The core is GNU AGPLv3; the ee/ directory uses a separate DayOtter Enterprise Edition License. 45

The project reaches beyond a meeting parser. The repository lists voice input, a voice receptionist, WhatsApp and SMS, running-late alerts, memory, and recaps. Those are ordinary assistant features individually; the interesting design choice is to keep an approval boundary around the irreversible one: changing someone’s schedule. 5

Try it. The developer quick start uses Node 20+, pnpm 10+, and Docker: clone the repository, run docker compose up -d, copy .env.example to .env, then run pnpm install, pnpm db:push, and pnpm dev. Open localhost:3000 after the dev server starts. 5

Caveat. The README calls the project early and not battle-tested at scale. It also lists enterprise SSO, SAML, SCIM, audit logs, SOC 2, and HIPAA support as not shipped. The open-core split matters if you need those controls rather than a personal or small-team scheduler. 5

What shipped. Prime Agent is an MIT-licensed "self-improving RLM agent" for coding workflows and long-running autonomous tasks. It uses a Recursive Language Model abstraction plus a Continual Harness for durable prompts, memories, skills, and subagent specifications. Its persistent Python control environment exposes IPython as a tool, lets the agent start subagents with rlm(...), and uses a daemon so work can be reattached later. 67

That is a different bet from a coding assistant that forgets everything when its terminal closes. Prime Agent makes the control environment part of the agent: the process can retain working artifacts, call subagents, and continue a task over a longer run. The payoff is persistence; the cost is that the agent has a much larger surface on which to make an expensive mistake. 7

Try it. The documented installer is curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh. Run prime-agent, then use /login on first launch. The repository is the place to inspect the available commands and the harness configuration before enabling autonomous work. 7

Caveat. The project warns that model-generated Python and project commands run with the user's permissions; workers and kernels are not a security sandbox. Budgets and gates reduce runaway behavior, but they do not turn an autonomous coding session into an isolated environment. 7

What shipped. Wallfacer is an MIT-licensed terminal session manager for Claude Code, Cursor CLI, Kiro CLI, and Codex. It collects sessions across agents and directories, then lets users name, tag, group, search, resume, or delete them from a full-screen terminal interface or the CLI. Its adapters are pluggable, and --json output makes the session catalogue scriptable. 8

Wallfacer is less ambitious than an agent runtime, which is exactly why it may be useful. The bottleneck for many coding agents is not generation; it is the half-finished thread buried in yesterday's terminal. A searchable, cross-agent index makes resuming the right context a small operational task instead of another round of archaeology.

Try it. On macOS or Linux, the README documents Homebrew installation with brew tap pradipta/wallfacer https://github.com/pradipta/wallfacer and brew install pradipta/wallfacer/wallfacer. Go users can run go install github.com/pradipta/wallfacer@latest; Go 1.22+ is required, and pre-built binaries are available from GitHub Releases. 8

Caveat. The current release page lists v1.2.3, while the Show HN post appeared on August 6. Wallfacer organizes and reopens sessions; it does not change the permissions or reliability of the underlying coding CLI. 910

What shipped. ModelFuzz is an MIT-licensed Python runtime guardrail for AI agents. Its @shield_tool decorator checks tool arguments before the wrapped function executes; a PolicyEngine applies ordered policies and raises ModelFuzzBlockError when a call is rejected. The README also describes a default-deny URL allowlist and a basic SensitiveDataFilter attached by the bare decorator. 1112

The placement matters. A prompt-injection attack can make a model produce an unsafe URL or argument even when the original task looked harmless. A check at the tool boundary gives the application one last deterministic place to reject that argument before a browser, shell, or API sees it. ModelFuzz is therefore a small control layer, not a claim that the model itself is trustworthy. 12

Try it. Install it with pip install modelfuzz or uv add modelfuzz; the optional scanner is available through pip install 'modelfuzz[scan]'. Start with a URL allowlist and a deliberately blocked test case before adding it to a real agent. 12

Caveat. The project is pre-1.0. Its own notes say that SensitiveDataFilter is a keyword tripwire, policies see one argument at a time, unrecognized argument types can pass through, and the runtime does not inspect prompts or model output. Treat it as one layer in a defense, not a complete prompt-injection solution. 12

What shipped. Nightcrawler is an MIT-licensed autonomous penetration-testing agent designed to run locally on a smartphone. Its repository uses the 1.2-billion-parameter LFM2.5-1.2B-Instruct-Heretic model through OpenCL on the phone GPU and reports testing on a OnePlus 8 with Kali NetHunter. It includes host discovery, service mapping, vulnerability finding, report generation, scope enforcement, a web dashboard, and dry-run mode. 1314

The novelty is the constraint: this is a security agent that does not need a cloud endpoint or a desktop GPU. That makes field use more plausible, but it also makes scope controls and authorization non-negotiable. A local model can reduce data exposure; it cannot make an unapproved scan legal.

Try it safely. The documented route is a Kali NetHunter chroot: run bash INSTALL.sh, check the service on localhost:8080, and use NC_DRY_RUN=1 python3 main.py before any live test. The project also documents a dashboard on port 8888 and a longer run script. Use it only against systems you own or have written permission to test, and keep the scope file narrower than the authorization. 14

Caveat. Compatibility is tied to the tested phone and Kali NetHunter environment, and pentesting results are not evidence that a target is safe. Start with dry-run mode and a deliberately disposable lab; do not interpret the project's autonomous workflow as permission to probe third-party infrastructure. 1314

Choose ADE if your problem is seeing several coding agents in one place. Choose Wallfacer if the problem is simply losing the right session. Prime Agent is the heavier experiment: it is for a task that needs continuity, subagents, and a process that can keep working after the initial prompt.

DayOtter is the most approachable test of approval-gated automation because a calendar gives you an immediate, visible feedback loop. ModelFuzz belongs lower in the stack, beside the tool definitions, where it can reject bad arguments before they become actions. Nightcrawler is the specialist pick: its value is on-device security work, and its first test should be a dry run in an authorized lab.

Taken together, these launches point to a practical split in indie agent building. The new work is less about giving an agent another personality and more about deciding where it runs, what it can touch, how it persists, and when a human must say yes. Those are the questions to answer before you let any of these tools near a real codebase, calendar, or network.