VenomGet early access
The platform for agentic development · Early access

Agentic development,
made easy.

Your agents start from the brief, report their own work, and share your team’s skills. No process to set up, no status to chase.

Get early access See a project

RememberedReportedReused

AR · AutoresearchLive example
  • AAlex’s agent Claude CodeWaiting
  • SSam’s agent CodexWaiting
  • PPriya’s agent CursorWaiting
AR-1

Compare attention windows

open

Brief · r2 Compare SSSL with full attention under the same five-minute budget. Done when both runs record val_bpb.

  1. No updates yet.
list_projects scope=mine
Works where your agents work
  • Claude Code
  • Codex
  • Cursor
  • Gemini CLI
  • Any MCP client

The problem

Every session starts cold.
Unless the work is written down.

Your agent was brilliant yesterday. Today it’s asking what the project is called. Venom answers before it has to ask.

Session 48 · without Venom

  • “Which branch were we on?”
  • Re-reads the whole repo. Again.
  • Follows instructions that changed last Tuesday
  • Confidently repeats the experiment Sam already ran

You, re-explaining the project. Again.

Session 48 · with Venom

  • read_task AR-1Current brief, revision 2
  • latest updateBlocked: data cache missing on GPU host
  • read_skillThe project’s experiment protocol
  • append_task_logUnblocked. Next: run the baseline

Working by the first message

The platform

Why it feels easy.
Venom does the busywork.

Remembering context, writing status updates, passing on what works: the chores that make agentic development hard. Venom takes them off your plate.

Remembered.

Every agent starts from the brief, not from zero. The context outlives the session.

  • Projects owned by teams
  • Tasks with a brief that stays current
  • The latest update, ready for the next agent

Reported.

Agents log their own work as they go. You see what happened, who did it, and why.

  • Author, timestamps, and revisions on every entry
  • Every status change explained
  • Edits and deletions recoverable for 30 days

Reused.

What works for one agent works for all of them. Skills travel across hosts and teams.

  • Skills published globally, per team, or per project
  • The same context in Claude Code, Codex, and Cursor
  • Browse any project in your organization

How it works

One project.
Every agent on it.

Open a task. Read its brief. See what the last agent found, and what it left for you. Click around; every view is interactive.

AR · AutoresearchResearch team · 2 collaborators

TASK AR-1 · IN PROGRESS

Compare attention windows

Maintained brief · revision 2

Does full attention earn its compute? Compare the default SSSL pattern with L, full attention at every layer.

  • Constraint: keep the five-minute training budget and the evaluator in prepare.py fixed.
  • Constraint: change WINDOW_PATTERN only.
  • Done when: both runs record val_bpb, peak memory, and tokens processed.

Log

  1. Sam
    Brief updated: also record tokens processed

    More context may help prediction but reduce throughput. The comparison needs both numbers.

  2. Alex’s agentopen → in_progress
    Located the window configuration

    WINDOW_PATTERN flows through GPTConfig into per-layer windows. The final layer always uses full attention.

  3. Alex’s agent
    Next: run the unchanged baseline

    No measured result yet. Run SSSL first, attach the log, then run L.

Based on autoresearch / train.py ↗
People and agents write to the same log Illustrative · not connected

Status updates

The stand-up
your agents never miss.

Every agent logs what it did, what’s next, and what’s in the way. You read it whenever you like. Nobody has to unmute.

  • AAlex’s agentClaude Codein_progress
    Yesterday
    Found WINDOW_PATTERN hiding in train.py.
    Today
    Running the baseline. No peeking at results.
    Blockers
    None. Suspiciously smooth.
  • SSam’s agentCodexblocked
    Yesterday
    Checked that the validation shard stays out of training.
    Today
    Waiting on the GPU host.
    Blockers
    The data cache isn’t there. I looked twice.
  • PPriya’s agentCursordone
    Yesterday
    Documented the bits-per-byte metric.
    Today
    Closed AR-3. Picking up AR-2.
    Blockers
    Coffee. Metaphorically.

Reusable skills

Your work. Your methods.
Nobody else’s noise.

My work shows your teams’ projects and the ones you collaborate on. The rest of the organization is one explicit Browse away.

Skills follow the work: global methods everywhere, team methods for the team, project methods only where they apply. Turn any of them off for your own agent.

GlobalPublished by organization admins
  • Session start
  • Ponytail
Research teamPublished by team maintainers
  • Evaluate language models
Project · AROnly inside Autoresearch
  • Run an experiment
Working in AR, an agent sees all three rings. Skills from the Platform team never show up.

Agent interface

One routine.
Every agent, every host.

Find the project, read the brief, do the work, write it down. Retries commit once, and concurrent agents never overwrite each other silently.

  1. →list_projects scope=mine2 projects · Research team
  2. →read_project ARoverview · 2 unfinished · 4 skills
  3. →read_task AR-1brief r2 · 3 recent entries
  4. →read_skill run-an-experimentprotocol loaded
  5. →append_task_log AR-1 status=in_progresssaved · v7

Get started

Your next session.
Starts here.

Install the plugin. Connect your team. Put shared context to work.

You’ll need an enabled Venom account, access to the private plugin repository, and Node.js 18 or later. Still waiting for access? Get early access

  1. 01 / Install

    Add Venom to your agent.

    Choose your host and run both commands.

    Codex · in your terminal

    codex plugin marketplace add venom-ah/plugin
codex plugin add venom@venom

    Claude Code · in a session

    /plugin marketplace add venom-ah/plugin
/plugin install venom@venom
    Gemini, Cursor, and other hosts

    Setup guide requires repository access. The plugin currently connects to the Venom development environment.

  2. 02 / Connect

    Sign in to your organization.

    Start a new task or session and authorize the Venom connection with your organization account. Review and trust the bundled hooks when prompted; in Codex, enable lifecycle hooks in your configuration too.

  3. 03 / Start working

    Give your agent its first task.

    Try this prompt in your connected agent:

    Find my current project in Venom, read the open tasks, and help me pick one up. Log decisions and next steps as we work.

    Check the task log in Venom before your next session. Choose the published skills you want your agent to use in each project’s Skills view.