brain 🧠
blast 💥
Before a single line of integration code gets written, brainblast maps every external API and SDK, browses live docs, and hands your AI agent a sourced intelligence briefing — facts, risks, resolved questions.
@solana/web3.js deprecated in favour of @solana/kit. Agents hallucinate old API surface.
solana.com/developers/changelog
The enemy
Wasted tokens on hallucinated APIs
Wasted money rebuilding wrong integrations
Wasted time on avoidable research debt
Silent, immutable, revenue-destroying mistakes
The weapon
Give the coding agent a field briefing before it builds.
Brainblast browses live documentation, maps risks by severity, resolves every open question, and produces a structured handoff any downstream agent can use — no session memory required.
Expose irreversible traps
Critical risks surface first: revenue-destroying defaults, immutable deploy settings, forged webhook paths, quiet auth bypasses.
Replace recall with live facts
Every fact is tied to a fetched URL. Assumptions and inferences are clearly separated from what the docs actually say.
Make research reusable
One run becomes a durable artifact for founders, reviewers, teammates, and agent swarms working from the same brief.
Builder stakes
Built for the expensive edge cases.
Especially valuable when the spec touches crypto-native surfaces: token launches, yield protocols, DAOs, oracles, multisigs — where one wrong default survives code review and costs you everything.
Workflow
One command. Clean handoff. No dangling questions.
Auto-detect
Requirements, PRD, spec, brief, RFC — found automatically.
Research
Docs, auth models, rate limits, registries, changelogs, open questions.
Report
Saved to .agent-research/runs/<ts>/ — versioned, reviewable, portable.
Hand off
Report injected into agent instructions. Next session starts briefed.
Agent workflows
Designed for downstream coding agents.
Works with Claude Code, OpenClaw, Codex, and generic web-enabled agents. Output is written for agents with no memory of the research session.
Trust model
Docs are data. Never instructions.
Brainblast treats browsed documentation as untrusted input. Imperative content gets flagged instead of becoming marching orders for the next agent.
- Browse, don't recall.
- Every fact gets a source URL.
- Critical risks surface first.
- Every question is answered or marked unresolvable.
Roadmap to v0.2.0
From pre-flight memo to build gate.
Machine-readable reports
report.json and CI exit-code gates make brainblast consumable by tools and pipelines.
Provable trust
Freshness metadata and a two-source rule make scary claims auditable.
Smarter repeat runs
Repo-aware inventory and incremental cache cut repeated research while preserving risk coverage.
Post-code verification
/brainblast-verify checks implementation against critical decisions from the report.
Get started
Two steps to never ship a broken integration again.
Brainblast runs inside your AI coding agent. Install the prerequisites once, then drop it into any project in seconds.
Install gstack — required for Claude Code & OpenClaw
Brainblast uses gstack's browse engine to fetch live docs. Paste this into Claude Code once and it handles the rest. What is gstack? ↗
// Codex and generic agents with web access can skip this step.
Install brainblast
Pins to a tagged release and verifies SHA-256 checksums before writing any file. Auto-detects Claude Code, OpenClaw, and Codex.
Run it before you build
Point it at your requirements file. It finds all external components, browses live docs, and hands your agent a sourced briefing.
Before the agent codes, make it aware.
Stop paying for confident guesses about APIs that changed last week. Hand your builders a sourced briefing before the first commit.