What We're Building Next
September 2026 · by the Merciful team
In our recent post on Vesa and the agentic loop, we shared how giving an AI model hands in a terminal changed how we maintain our own systems. That post prompted a fair question: if tool-wielding loops are so effective behind the scenes, when do those capabilities reach the actual chat interface?
Roadmap posts often turn into marketing brochures full of speculative promises. We want to avoid that. This is an honest summary of what is currently on our workbench, why we are building it, and the technical realities of getting it right.
The Problem: The Limits of the Raw Text Box
Right now, Merciful is a clean, responsive conversation box. For drafting, brainstorming, or talking through a problem, that works well. But as soon as you rely on the model for day-to-day work, the boundaries of raw chat become obvious:
- Zero persistent context: Every room starts completely blank. If you want the model to write in a specific style, act as a strict code reviewer, or keep answers concise, you have to re-explain yourself every time.
- Frozen knowledge: The model cannot check a current fact, consult live documentation, or verify whether an API has changed since its training cutoff.
- Friction with real files: Copying and pasting 300 lines of code or an entire document into a browser textarea is clumsy and destroys formatting.
- No execution feedback: The model can write code or draft HTML, but it cannot test whether it runs or see what it looks like.
The difference between a basic chatbot and an effective assistant is not a larger parameter count—it is tools. Here is how we are building those tools into Merciful.
1. Custom System Prompts & Work Baselines
Having to re-type preamble instructions at the start of every session is friction that shouldn't exist. We are adding system prompt configuration and reusable baselines (what we previously called "personas").
Rather than decorative avatars, these are explicit system instructions that shape how the model reasons, formats answers, and approaches tasks from token zero. Whether you need a concise technical editor, an interactive tutor, or a strict debugger, you can save and switch baselines without re-priming the room.
The main engineering consideration here is context window transparency: large system prompts consume input tokens before you even submit your first prompt. We are designing the interface to clearly indicate how much context a baseline uses so you never wonder why your token allowance moved.
2. Live Data Retrieval and Reference Grounding
When an LLM does not know something, standard autoregressive sampling often leads to plausible-sounding hallucinations. For technical work, that undermines trust immediately.
We are building retrieval tools directly into the runner pipeline. Starting with curated reference sources (such as Wikipedia and official software documentation), the model can formulate search queries, fetch real-time excerpts, and ground its answers with explicit source citations.
Retrieval naturally introduces external network latency. Because we recently rolled out server-reported progress tracking and ETAs in the chat interface, the UI will clearly indicate when a model is searching external sources versus generating tokens.
3. Document & File Handling
Working with documents should not require juggling multiple browser tabs and pasting blocks of text back and forth. We are working on native file attachments that allow the model to ingest, parse, and reference whole files.
Beyond ingestion, we want the output to be practical. When you ask the model to review a script or edit an article, it should be able to produce clean unified diffs or a direct downloadable file rather than burying changes in a monospace chat block.
4. Open Image Generation
Visual generation has been one of the most consistent requests from our community. However, image diffusion models operate under fundamentally different economic and hardware constraints than text LLMs: generating a single high-resolution image can consume substantial GPU memory and compute time.
Many services address this by introducing confusing secondary currencies, "credits," or aggressive upsells. We are avoiding that. We are integrating open diffusion models directly into our existing cluster architecture and mapping them onto the same transparent daily token system. Free verified accounts will have a realistic daily allowance, and Pro subscribers will have higher capacity.
5. Looking Further: Sandboxed Environments
In our Vesa writeup, we described giving an agent shell access to inspect files and run diagnostic commands. Bringing that capability to users is the logical long-term destination, but it requires extreme care.
Allowing an AI to generate and execute code, or preview a rendered webpage, is not just about writing HTML—it requires hardened container isolation, rigorous resource limits, and network safeguards to prevent abuse. We are currently testing isolated web sandboxes so that code written in chat can be previewed safely in real time. Further down the line, we are exploring lightweight local desktop integration for developers who want the agentic loop operating on their own machine, without sending local files or shell access through remote servers.
How We Keep This Sustainable
Every new feature we add requires dedicated GPU hours and runner bandwidth. As we noted in our post on how Merciful runs, we do not monetize your chat logs, sell behavioral profiles, or insert sponsored links.
Our $5/month Pro subscription exists specifically to cover our hardware hosting bills. As these new capabilities roll out, our model remains straightforward: core features remain accessible to free verified accounts with practical daily quotas, while Pro accounts provide the higher limits that power users need and keep the service independently funded.
What We Are Still Not Building
It is just as important to be clear about what we will never pursue:
- Simulated companionship: No artificial emotional memory designed to foster dependency. Merciful is a tool designed to help you think and build, not a pretend relationship.
- In-stream advertising: No sponsored recommendations, affiliate injections, or banner clutter in your chat room.
- Data harvesting: Your room history is stored so you can revisit your past conversations, not to train ad-targeting models or sell data to third parties.
Ship It When It Works
We don't publish artificial release deadlines or announce features before the underlying engineering is viable. We are building these tools because they solve real limitations we experience every day while using our own software. As each feature reaches stability on our test runners, we will roll it out directly to your interface.
— The Merciful Team