Manual context curation for AI coding agents
Problem
Developers working with LLM coding agents feel they've lost control over what context goes into each prompt: context is auto-managed inside frontends like Claude Code, with no way to pick exact files, fork a session, or take asides/notes without disrupting the running task. An HN commenter describes increasingly wanting explicit control over context assembly and session forking.
Opportunity
A context-management layer/IDE companion for coding agents that gives developers explicit, diffable control over the files and discovery artifacts fed into each prompt, plus first-class session forking and side-channel notes.
Market analysis
Explicit context assembly is already a staked claim — Repo Prompt markets exactly this ('assemble the perfect context') and sits between local repos and CLI agents like Claude Code. The unserved slice is treating context as versioned, diffable artifacts plus first-class session forking, but the space moves brutally fast and harness vendors keep absorbing context features natively.
Market · Power users of CLI coding agents (Claude Code, Codex CLI, Gemini CLI) frustrated by opaque auto-context; active HN discussion and a fast-expanding tooling ecosystem.
Pricing · Developer tools in this niche run freemium desktop/CLI apps with Pro tiers in the $10-20/month range; open-source variants monetize hosted team features.
Pros
- + Sharp, articulate early-adopter audience that pays for leverage.
- + MVP is a local CLI/desktop tool with zero infra cost.
- + Session forking and side-channel notes are genuinely underserved.
Cons
- − Repo Prompt and adjacent tools already own the context-assembly positioning.
- − Agent harnesses ship native context improvements quarterly, shrinking third-party surface area.
- − This audience churns to whichever tool is newest — loyalty is near zero.
Existing / similar tools
Source
Hacker News (Ask HN)
The real differentiator is not file-picking — every context tool does file-picking — but treating context as code: versioned manifests that diff cleanly, get reviewed like PRs, and travel with the repo so the same context set replays months later or on a teammate’s machine. That reframes the product from a UI convenience into session infrastructure, which is stickier and harder for harness vendors to absorb (they optimize for zero-config magic, and magic is the thing this audience explicitly wants to opt out of). The brutal part is timing: every quarter of native improvements shrinks the gap a third-party layer can fill, so the defensible window is the diffable-manifest angle, not the context-selector angle. Build the former or don’t build.