Park an idea.
Add it to the board yourself, even if it needs more thought before an agent starts.
Open-source agent development environment
One prompt. A whole software factory.
Describe the outcome once. A coordinator agent plans the phases, puts each task on the model that fits it, and checks in with you before anything lands.
Podium runs on your computer or a VPS. Install it there, then follow and steer your agents from this phone's browser.
A native iPhone app is in private beta.
Downloading PodiumOpen it and drag Podium to Applications. Try again
Open source. Run locally or on your own VPS.
Coordinator
ChatCLI
04 · SHIP
You stay the captain.
Illustrative Kelvra launch walkthrough
- 01
Describe the outcome.
KEL-2 holds the whole job. - 02
The coordinator plans it.
Four phases, eleven tasks, each with what it waits on. Research goes to Sonnet, design to Opus, the build to GPT-6.1 Sol.
KEL-12 waits for KEL-9 and KEL-10. - 03
Agents work in parallel.
Every task gets its own worktree and agent. Finished work stays on the deck, checked off, until you hide it.
KEL-13 starts when the design lands. - 04
You make the call.
An Opus reviewer checks the table and toggle. Then the coordinator brings you the draft with its evidence before anything merges.
KEL-18 waits for your answer in KEL-17.
Works with your coding agents
- Claude Code
- Codex
- Cursor
- OpenCode
- Grok
- Pi
An agent-native issue system, that is also yours
Put an idea on the task board before you are ready to discuss it. Or start in chat: tell an agent the idea and it creates and describes a task for you. You and your agents use the same board to shape ideas, organize work, and follow each task through review.
Scroll across to explore the full view.
Scroll across to explore the full view.
Scroll across to explore the full view.
Turn discussion into tasks.
Start in chat instead. Ask an agent to write the task, split the work, and link dependencies.
Let agents propose what is next.
While working, agents can raise follow-up ideas for you to accept or defer.
Keep results with the work.
Plans, screenshots, recordings, and other artifacts stay on the task beside its discussion.
Any agent can become your coordinator
Ask the agent you're working with to coordinate the job. It can create subtasks and dependencies, bring in other agents, and follow their results. You choose who does what and can change the plan as the work develops.
You could ask
"Build the Kelvra vs. the rest comparison page for launch week. Research the three closest rivals with Sonnet 5.5, design with Opus 5.5, build with GPT-6.1 Sol. Review code where it matters, and check in with me before it ships."
- 01Research, in order
Sonnet 5.5 builds the rival pricing matrix in KEL-8. The claims audit and sentiment scan in KEL-9 and KEL-10 start from it. - 02Design from the research Opus 5.5 lays out the comparison table in KEL-11 and writes the hero and claims copy in KEL-12.
- 03Build on the model that fits Codex on GPT-6.1 Sol builds the table, the plan toggle and the schema markup in KEL-13 to KEL-15.
- 04Review, then your call An Opus reviewer checks the table and toggle in KEL-16. The coordinator asks you in KEL-17 before KEL-18 lands on main.
Tasks, decisions, dependencies, handoffs, and results are recorded as structured state in Podium's issue system. Your software factory takes shape without a rigid process to define up front.
Your Software Factory at a glance
See which agent is on each task, what waits on what, and how far the work has come. Open any task to reach its agent and conversation.
Step in only when it counts.
Agents work on their own. The deck shows where everything stands, so you can follow along and step in if you want to.
Spot the bottleneck.
See what waits on what, and which task unblocks the rest.
Nothing gets lost.
Done work stays in place, one click from its conversation.
Small task or big job.
When you run a few tasks yourself, the deck shows what needs you next. On a bigger job, the coordinator handles those questions and checks in with you on the calls that are yours.
You stay the captain
Discuss an approach, challenge a decision, inspect an implementation, or change direction as the work develops. Work through the coordinator or engage directly with an individual agent. Chat and CLI views give you different ways to participate in the same work.
Coordinator
ChatCLI
Before you start the SEO task, what is the plan for pricing? I don't want the two builders fighting over the same files.
The pricing cards KEL-14 and SEO markup KEL-15 run in parallel on Codex GPT-6.1 Sol, both from the merged table KEL-13.
The pricing builder owns compare.css. The SEO builder owns only the page <head>, so the two can't conflict.
KEL-15 asks whether search data prices in USDReply
Working0:42
Coordinator
opus 5.5 · highChatCLI
> Before you start the SEO task, what is the plan for pricing? I don't want the two builders fighting over the same files. ● The pricing cards (KEL-14) and SEO markup (KEL-15) run in parallel on Codex GPT-6.1 Sol, both from the merged table (KEL-13). ● Bash(git diff main --stat) ⎿ 6 files changed ● Read(compare.css) ⎿ Read 188 lines ● Bash(bun run check) ⎿ lean gate green in 48s ● The pricing builder owns compare.css. The SEO builder owns only the page <head>, so the two can't conflict. ● Bash(podium issue mail send KEL-15) ⎿ Sent ✻ Working 0:42
❯
⏵⏵ auto mode on (shift+tab to cycle)
Desktop, server, phone
Give your agents a machine of their own.
Start on your Mac with the coding agents you already use. When agents should keep working after your laptop closes, run Podium on a Linux VPS you own. Every Podium app connects to every Podium server.
Desktop app
Orchestrate agents running on your Mac, or connect to a Podium server on your VPS.
Download for MacServer
Run agents on your VPS, always on. Control them from the desktop app, your phone, or any browser.
your-vps$ curl -fsSL https://podium.do/install.sh | shPhone
Control agents on the go. Connect to Podium on your own computer or a VPS.
Use Podium on your phone
Review and workspace tools
Podium keeps adding tools that make agent work easier to coordinate and easier for you to review. These help with decisions, long documents, task evidence, and understanding usage.
Agent offers
Agents return with a clear decision and the evidence to make it. Approve from the card, or request changes.
Needs your decision · 1mComparison page draft ready for your check-in
Built and reviewed on this issue's branch, with pricing sourced and dated. To judge: does the table read fairly against all three rivals? Approving lands it on main.
Calm reading mode
The document takes a full-height column over the dimmed app. The outline sits in the left margin, the time left in the right. Scroll this one.
Readingreview-compare.md
Review: comparison table and plan toggle
Scope
src/pages/compare.html and src/styles/compare.css, checked against the design spec, the approved copy and data/rivals.json. Base c9ec3f0, with KEL-13 and KEL-14 merged.
Method
Every figure recomputed from the JSON by hand. The page rendered in headless Chromium at eight widths from 320 to 1280px, with keyboard-only and JS-disabled runs. Contrast computed from the rendered colours of every visible text node in main.
Result
One defect fixed, a claim that went beyond the data. No numeric errors. Five low or subjective items left for the coordinator.
Checks
| Check | Result |
|---|---|
| Numbers match JSON | Pass, all 16 card figures and 4 price-row ranges |
| Table semantics | Pass, one table with a caption and scoped headers |
| Capability cells | Pass, all 32 cells |
Left for the coordinator
Both "Create your free booking page" buttons link to the same placeholder as the homepage, because there is no signup URL yet. The footnote on rival pricing should carry its review date. Whether "fairest" in the hero overreaches is a judgment call for you.
The proof stays with the task
Every artifact a task collects gets one page: screenshots, recordings, documents and reports, filterable and searchable. It stays reachable after the original worktree is gone.
Kelvra coordinatorKEL-2 artifacts
Screens 4
Draft comparison page, desktopImage · Oct 31
Draft comparison page, phoneImage · Oct 31
Pricing cards at 1280pxImage · Oct 30
Plan toggle walkthroughVideo · Oct 30
Documents & reports 5
comparison-draft.htmlHTMLOct 31
compare-spec.mdDocumentOct 29
rivals.jsonDataOct 28
review-compare.mdReportOct 31
Pricing claims auditReportOct 28
Proposed tasks
Follow-up work does not vanish into a chat or quietly expand the current job. Agents file it on the deck, and you start it, backlog it or dismiss it.
KEL-18Land on mainDone
KEL-19Launch-week changelogWorking
Proposed3
KEL-24Calendrix switcher guideBacklogStart
KEL-25Link the table from the pricing pageBacklogStart
KEL-26Refresh rival prices every quarterBacklogStart
Know what the work cost
Recorded token usage and API-equivalent cost estimates, for the task and its sub-tasks.
COST≈$11
$10.76
at list price for the same tokens, not what you were billed
Where it went
This task $4.81 16 sub-tasks $5.95
claude-opus-5-5$8.53
gpt-6.1-sol$1.41
claude-sonnet-5-5$0.82
10 sessions · 258 replies24.0M tok
Illustrative usage; costs are API-equivalent estimates.
Put your next
idea to work.
Bring your coding agents into Podium and start building.
