I build AI agent infrastructure — agent runtimes, browser-driving agents, provider routing and MCP tooling. I also ship full-stack product. India-based, working in public.

Patches to steel-dev/steel-browser — the open-source browser API for AI agents (7.5k★):

#353 — session timeout with automatic release. The SDKs exposed a timeout on session creation; the self-hosted server accepted it, always reported timeout: 0, and kept the session live until released by hand. Implemented server-side release, guarded the timer so a stale one can't release a session that already got replaced, and fixed a spread-order bug in resetSessionInfo that was silently discarding caller-supplied values. 5 unit tests.

#354 — DevTools inspector under USE_SSL. .replace("ws:", "") no-ops on wss:, so the frontend was handed a malformed URL and CSP-blocked. Stripping the scheme alone isn't enough: the DevTools frontend derives its protocol from which query parameter carries the value, so a secure endpoint sent as ws opens an insecure socket and gets blocked as mixed content. Switched the parameter too, leaving non-SSL output byte-identical. 6 unit tests.

#355 — session viewer rendered the wrong session. No Router was ever mounted, so useParams() returned undefined and every viewer URL fell back to sessions[0]. The header meanwhile hand-parsed window.location, so the header named one session while the body rendered another. Measured before and after in a real browser: 3 of 5 API calls hit the list endpoint before, 0 of 1 after.

browserless/browserless#5551 — /json/new now opens the requested URL. The endpoint ignored its request and hardcoded about:blank; the actual page is only created later when the client connects to the returned websocket. Preserved the raw DevTools query form before Browserless normalizes it, carried the target across that boundary, declared it in the strict websocket schema, and applied the navigation-security policy at both entry points. 69 core tests, 11 JSON-route tests, and a real Chromium end-to-end check.

XOMNI — one agent, every feature, every free model. Composes 7 open-source coding agents on one host with 35 plugins, a provider pool routing 25 verified free models, and a 1,251-test suite.

unified-agent — Hermes host + 13 modules: context compaction, sandbox gating, MCP catalog, repo mapping, GitHub ops, local model serving, verification runner.

StrawManus — a browser-control agent fusing Manus's flat 12-tool action layer with a session, natural-language selector and skill-reuse layer. Runs in your real logged-in browser, self-heals when sites change, replays successful tasks as skills. Ships an MCP server.

BX Browser — DOM-first, vision-capable browser agent. Reads the accessibility tree for 2–4s steps and falls back to vision only when the page demands it. Session keep-alive, model-ladder fallback, zero-dependency core.

OpenBerry — a local agentic browser for Windows. An agent drives your real Chrome/Edge across tabs over CDP, logged in as you. 100% local, MIT.

hermes-skills — custom skills for Hermes Agent: delegation recovery, session artifact recovery, CLI APK builds, constrained-budget scraping.

VIGIL — autonomous trading-ops agent. Scans markets around the clock, rule-checks, signals, monitors, journals. Google "All Things Agentic".

PantryWatch — stops food-bank donations expiring. AWS Strands Agents SDK.

Hackathon Radar — Devpost scanners, prize and deadline tracking, and win-pattern analysis. Most hackathon losses are selection losses, not build losses.

StackPilot — Android offer console for Indian delivery riders. Allowlisted read-only notification capture, deterministic on-device scoring and trip-chain routing. Kotlin, R8, signed releases.

Python TypeScript Next.js React Kotlin Node

Agent runtimes · MCP servers · Chrome DevTools Protocol · Playwright · multi-provider LLM routing · Cloudflare Workers · AWS Strands · Vercel

📫 painpratham@gmail.com