Skip to content
Himadri MishraAI engineer, 8+ years · Remote, India

I build agents, and the MCP servers, CLIs and SDKs they run on.

Himadri Mishra

I build AI products at Mudita Studios and write Agent Experience, a guide to how agents find, choose and use software. Before that I was at Firecrawl (YC S22), and before agents I did production ML, starting in 2018.

The last big thing I shipped: Firecrawl's hosted MCP server, rebuilt so sign-in works and agents load half the tool-metadata tokens with no drop in task success, in a paired A/B test. Firecrawl's Claude and Codex connectors run on it.

See selected work
01 · Mudita Studios

A coding agent for Slack and Jira requests

A ticket in, a draft pull request out, checked outside the sandbox.

Read more02 · Firecrawl

The hosted MCP server, rebuilt

Sign-in that works, and agents load half the tokens.

Read more03 · Firecrawl

The Agent Experience harness

Real coding agents test every release before it ships.

Read more
01Selected work

What I built.

Mudita Studios · since May 2026

A coding agent that answers Slack and Jira requests with pull requests

Someone asks for a change in Slack or Jira. The agent reads the ticket, works in a fresh sandbox with no credentials, writes the change and its tests, and runs them. A second model that did not write the code reviews the diff and can send it back. Only a patch leaves the sandbox.

  • Then the checks the agent cannot touch: the project's tests run again outside the sandbox, pages are checked in a real browser, and a new test must fail on the old code first.
  • The bot cannot merge. A person reviews the draft, and a reply in the same thread updates the same pull request.
Inside the coding agent
Inside the coding agentA request in Slack starts a coding agent in a sandbox with no credentials; a second model reviews its diff and can send it back; only the patch leaves; checks outside the agent's reach run before a draft pull request goes to a person, whose reply sends the work back.SANDBOX · NO KEYSRequestSlack · Jira ticketCoding agentexplores, codes, testsSecond modelreviews the diffIndependent checkstests · browser · proofDraft pull requestbot cannot mergePerson reviewsmerges or repliesREVISEPATCH ONLYPASSREPLY · SAME PR
The agent cannot commit or push, and each new test must fail on the old code before a pull request opens.
Firecrawl · July 2026

Firecrawl's hosted MCP server, rebuilt

Before and after
The hosted MCP server, before and afterBefore, an agent reached one server with an API key in the URL and an OAuth flow that was mis-wired and often failed. After, an agent signs in with Firecrawl OAuth to an account endpoint, or uses a keyless endpoint that needs no account.BEFOREAgentany clientOne serverkey in the URLOAuth mis-wiredAFTERAgentany clientAccount endpointFirecrawl OAuthKeyless endpointno account needed

The old server took the API key in the URL, and its OAuth flow was wired wrong, so sign-in failed a lot. I rebuilt it on Firecrawl's OAuth with two endpoints: one for accounts and one that needs none, so an agent can start without signing up and is told what to do when a tool needs an account.

Agents now load half the tool-metadata tokens, and task success stayed the same in a paired A/B test. Firecrawl's Claude and Codex connectors run on it.

Firecrawl · May to September 2026

The Agent Experience harness

Inside the harness
Inside the Agent Experience harnessA loop of six steps, pin the builds, run real agents, keep every trace, check then judge, report findings and fix the product, then back to the first step to re-test; new releases enter at the first step, and search checks without agents feed the findings.01Pin the builds02Run real agents03Keep every trace04Check, then judge05Report findings06Fix the productSearch checks, no agentsNEW RELEASESRE-TEST
Agents: Claude Code, Codex, Cursor and OpenCode, in sandboxes.

How Firecrawl found out whether agents could use its tools. Before a release shipped, real coding agents (Claude Code, Codex, Cursor, OpenCode) ran the candidate build against the published one in sandboxes set up like a user's, in a paired A/B test.

  • Every tool call, page read and answer was kept raw, and every quote in a weekly report was checked against its run.
  • What the runs found went into the docs, API, MCP server and website, and the next run tested it.
Also
  1. Mudita StudiosThe research product. Main engineer on a product whose answers cite their sources: data connectors, the test harness and release gates. I also set up CI/CD, previews and daily browser smoke tests for a second product.
  2. Firecrawl · launchesThe paper index. With the research-index skill, agents reached Firecrawl's new biomedical paper index in all 48 test runs. Without it, none did. skills #10
  3. Firecrawl · tool callsSearch routing. With routing instructions in the server, every Claude Code run and two thirds of Codex runs used Firecrawl's search instead of the built-in one, over 15 runs each. firecrawl-mcp-server #240
  4. Firecrawl · recoveryErrors an agent can act on. A spent limit or a wrong key now returns what to do next instead of ending the task, and API errors come back as JSON, not HTML pages. firecrawl-mcp-server #365 firecrawl #4211 firecrawl #4552
Before agentsFull resume
  1. 2025 to 2026Knit, Senior AI Engineer. I designed the AI pipeline of its research platform; reports that took 2-3 days now take under an hour.
  2. 2023 to 2024Epic! for Kids, Senior Research Engineer. Search and recommendations for a children's reading app, and their ML platform.
  3. 2018 to 2023Osmo and Whodat, Computer vision: worksheet recognition for a kids' learning product, and feature detection for augmented reality.
  4. 2013 to 2018IIT (BHU) Varanasi, Computer Science. SN Bose Scholar, with a research internship at UC Berkeley under Professor Dawn Song.
02Latest

What I have written lately.

Agent Experience is my guide to how agents use software, with an open readiness rubric, an evidence register, a checker and a public MCP server.

  1. 12 Oct 2026What does an AX engineer do? Guide
  2. 11 Oct 2026Five meanings of agent experience Guide
  3. 09 Oct 2026Agent readiness scores compared Guide
  4. 08 Oct 2026What changed in AX, edition 1 Digest
  5. 30 Sep 2026How to expose a 3,000-operation API to agents Guide
03Built on the side

Built on the side.

01Talkies2026

Dictation for Linux: hold a key, speak, and the text is pasted at your cursor. Speech is transcribed on your own machine.

02Qwen3.6-35B on an 8 GB laptop GPU2026

A 35B model on a laptop GPU at 43 tokens per second with 128K context, with launch scripts, a tuning guide and a coding benchmark.

03ubuntu-maclike-touchpad2026

Mac-style touchpad gestures on Ubuntu: three-finger drag, four-finger workspace switching, pinch zoom.

04laptop-powersave2026

Battery modes for a Linux laptop that doubles as a home server: it caps the CPU when the charger is unplugged and powers down the idle NVIDIA GPU without a reboot.

05cctv-edge2026

A local web console for a Dahua-compatible NVR: live view, Frigate detection and archive backfill, nothing exposed to the internet.

Earlier open source
01NTM-One-Shot-TF238 stars

One-shot learning with memory-augmented neural networks, in TensorFlow.

02Botnet-Detection-using-Machine-Learning178 stars

Botnet detection with classical machine learning, my bachelor's project.

03Parallel-Youtube-Titles-to-MP314 stars

Search YouTube by title and save the audio as tagged MP3s, in parallel.