Open-source infrastructure for software engineering with AI agents
Agents come and go. The project stays.
Charrette gives a software project a memory it owns and a coordinator that moves the work between AI agents. Any agent, any model, your own subscriptions, on your own servers.
Scroll ↓
Every major AI coding company is building a project layer, and each one ties it to the agent it sells. Charrette keeps project knowledge open and independent, and uses it to run work across whichever agents are best this month.
The problem · Memory
The project has no memory of its own.
Why a decision was made, what was already tried, which rule came from a customer. That knowledge is spread across tools and people’s heads. Change agents and most of it stays behind.
Charrette’s answer → a memory the project owns
Claude Codesix months of work
→Codexday one
- CLAUDE.md
- AGENTS.md
- A chat from March
- .cursor/rules
- PR #1147 review thread
- What Priya remembers
- docs/architecture.md
- A compacted chat summary
- Slack, #payments
- Why we don’t use cascades
Ten places this project’s knowledge lives today.
A better model ships. The team switches to Codex.
2 of 10 come along. The rest stay behind.
Codex, an hour later“Should I add a cascade delete here?”
The problem · Attention
The developer is the message bus.
You know the steps a fix should go through. Run them by hand and every hand-off waits for you to notice, and every step depends on you remembering it.
- minutes the agents sat waiting for you
- 62
- results you carried from one agent to the next
- 5
- steps that never ran
- 2/ 7
- 01Triage Claude CodePaste the ticket in. Explain how sessions work, again.09:14 · waited 8 min
- 02Implement Claude CodeOpen Codex. Paste the diff. Explain the ticket, again.09:49 · waited 15 min
- 03Security auditNobodyNobody remembered that token rotation is a sensitive path.Never ran
- 04Review CodexCopy the findings back into Claude Code.10:24 · waited 9 min2 of 3 findings passed on
- 05Repair Claude CodeBack from stand-up. Ask Codex to look again.10:53 · waited 23 min
- 06Re-review Codex“Looks clean.” Merge it before lunch.11:30 · waited 7 min
- 07Acceptance testNobodyThe criteria were in the first chat, three hours back.Never ran
Merged at 11:40. Two steps never ran and one finding was lost in a paste, so the bug ships at one call site, and so does an unaudited change to token rotation.
With Charrette → all seven steps, every time, and one question for you
The answer · The coordinator
The graph isn’t drawn in advance. Each result decides the next step.
Task 418After a user’s role changes, their old permissions keep being served until the session token refreshes.
The answer · Project memory
Every task leaves the project knowing more.
payments-serviceWhat the project knows, and how sure it is of each thing.
09/ 9
Record
Coordinator · writes back to project memory
Three entries go back to the project: a new convention, your decision, and the unresolved refresh window. The next task inherits all three, whichever agent runs it.Back on the main line, once everything the evidence added is settled.
5people, one record
Shared by the team
next · with Charrette for teams
Everyone on the project reads and writes the same memory, with authors and history. A call Priya made in March briefs the agents of someone who joined last week.The project remembers, so no one person has to.
- 01TriageCoordinator · your choice of model
It starts from project memory, not a blank prompt. It reproduces the bug, writes three acceptance criteria and attaches the six facts that apply. Two of them disagree about the refresh window, so both go along, marked disputed.
- 02ImplementClaude Code · paid by your Claude plan
Routed on evidence: this agent has the best record on this module’s last five tasks. It gets the criteria and the failing test, not a summary of them, and fixes the bug by rotating the session token when a role changes.
- 03Added · sensitive pathSecurity auditOpen-weight model · on your own GPUs
Runs on your own hardware, so this code never leaves your network, and it is a third model family. One finding it can’t settle on its own: if rotation fails, the request now fails too.
- 04Added · your callDecideYou · one question, one tap
Fail the request, or retry the rotation once? It changes what users see, so it goes to a person, with the proposal from project memory attached. The review doesn’t wait for your answer.
- 05ReviewCodex · a different lab from the author
Session code gets a review from a lab other than the author’s. It finds the same stale read at three more call sites.
- 06Added · 3 call sitesRepairClaude Code · paid by your Claude plan
It gets all three findings and your answer on rotation, with the original criteria, so it can’t fix one problem by reintroducing another.
- 07Added · every repairRe-reviewCodex · checks the repair against the findings
Checked against the three findings, not treated as a fresh review. All three call sites are fixed. Clean.
- 08Added · user-facingAcceptance testGemini CLI · a fresh agent that never saw the code
The three criteria from triage, checked on staging: a demoted user loses access on their next request, other sessions are untouched, nobody is logged out. All three pass, and so do 412 tests.
- 09RecordCoordinator · writes back to project memory
Three entries go back to the project: a new convention, your decision, and the unresolved refresh window. The next task inherits all three, whichever agent runs it.
The answer · Project memory
Every task leaves the project knowing more.
Task 418 started from six facts and wrote three back. They join the 214 facts payments-service already holds, each with a status and a source. When two disagree, both stay until a person decides.
- New · SettledRole checks read live permissionsconvention · task 418
- New · SettledA failed rotation retries onceyour decision · task 418
- New · ProposedRefresh window: 15 or 5 min?open question · task 418
- SettledToken rotation is sensitivesecurity policy
- SettledCross-lab review on session codeproject policy
- SettledWebhook deliveries are idempotenttask 407
- SettledMoney is integer minor unitsPriya · March
- Settled · disputedSessions refresh every 15 minreview · task 391
- Seen once · disputedProduction reads 5 minseen once · task 402
- Seen onceClaude Code leads on this modulelast five tasks
- No longer truePermissions are cached per sessionretired by task 418
- ProposedWrites fail fast; deliveries retryproposed · task 425
- No longer trueRefunds are synchronousno longer true · task 388
Next, with Charrette for teams: 5 people on one record. A call Priya made in March briefs the agents of someone who joined last week.
What orchestration actually means
Launching a sub-agent is delegation. Orchestration is deciding what happens next.
- Delegation
One agent starts another and waits for text back.
Most agent tools - Parallel agents
Many sessions at once. You carry the context between them.
Most agent tools - Fixed pipeline
The same steps every time, whether they’re needed or not.
Workflow features, CI - Supervision from evidence
Each result decides the next step: a finding adds a repair, a sensitive file adds an audit. You’re asked only when it matters.
Charrette
The vendors sell steps 1 to 3. Step 4 only works as one thing: the graph, the routing, the person and the memory.
The experience
Run agents the way a good lead runs a team.
- 1
A stand-up, not a chat log. What moved, what’s running and what’s waiting on you, as a report or as a board.
- 2
Only decisions reach you. Product calls, irreversible actions and contradictions in memory.
- 3
The right agent for each step, with review from a different lab from the author’s.
- 4
Built as an instrument. Calm, dense and keyboard-first, for hours of daily use.
1Since you left yesterday at 18:10
Moved
- ✓Refund idempotency keysPR #1184 merged
- ✓Checkout timeoutscause found: pool exhausted under retry storms
- ✓Webhook retry contractpromoted to settled memory
Needs you 2
Task 418 · a decision
If token rotation fails, should the request fail, or retry once?
Found by the security audit. No decision on record; memory proposes retrying once, from task 402.
Running 3
- 418Stale permissions after a role changeReview · Codex
- 419Migrate billing webhooks to v2Implement · Claude Code
- 421Drop legacy_sessions tableReview · Codex
- 424Currency rounding in refund totalsTriage · Coordinator
CoordinatorI added a security audit to 418: the diff touches token rotation, which a rule from 4 February marks sensitive.
Needs you 12
Running 43
Held 1
Settled 31
Where Charrette sits
A persistent layer between your team and whichever agents you use.
AgentsReplaceable. Use whichever is best this month.
Yours
Mixing labs is the point: a reviewer from a different lab from the author is one of the strongest checks there is, and no lab will sell you an independent reviewer of its own model.
Why now · The frontierOpen weights
The best model now changes every few weeks.
Each bar is the model at the top of the Artificial Analysis Intelligence Index, from its release until a better one shipped. The lead changed twice in 2023 and eleven times in 2025. Four labs have held it since 2024.
And one you can run yourself ships every two weeks.
Each dot is an open-weights model at least as good as the top model of six months before: one a year until 2025, then eight, then 19 so far this year from seven labs. The best, MiMo-V2.6-Pro, matches the frontier of June.
- 2023GPT-3.5 TurboGPT-3.5 Turbo OpenAI · 30 Nov 2022 · 5.5GPT-4GPT-4 OpenAI · 14 Mar 2023 · 6.7GPT-4 TurboGPT-4 Turbo OpenAI · 6 Nov 2023 · 7.0Llama 2 Meta · 18 Jul 2023 · 5.7
2 new #1 modelstop score +1.5
1 open, close behind
- 2024GPT-4 TurboGPT-4 Turbo OpenAI · 6 Nov 2023 · 7.0Claude 3 OpusClaude 3 Opus Anthropic · 4 Mar 2024 · 8.7o1-previewo1-preview OpenAI · 12 Sep 2024 · 11.4o1o1 OpenAI · 5 Dec 2024 · 15.2Llama 3.1 405B Meta · 23 Jul 2024 · 7.3
3 new #1 modelstop score +8.2
1 open, close behind
- 2025o1o1 OpenAI · 5 Dec 2024 · 15.2Claude 3.7 SonnetClaude 3.7 Sonnet Anthropic · 24 Feb 2025 · 17.7o3o3 OpenAI · 16 Apr 2025 · 20.2Claude 4 OpusClaude 4 Opus Anthropic · 22 May 2025 · 20.6o3-proo3-pro OpenAI · 10 Jun 2025 · 21.9xGrok 4Grok 4 xAI · 10 Jul 2025 · 22.5Claude 4.1 OpusClaude 4.1 Opus Anthropic · 5 Aug 2025 · 22.8GPT-5GPT-5 OpenAI · 7 Aug 2025 · 23.0GPT-5 CodexGPT-5 Codex OpenAI · 23 Sep 2025 · 24.9Gemini 3 ProGemini 3 Pro Google · 18 Nov 2025 · 28.0Claude Opus 4.5Claude Opus 4.5 Anthropic · 24 Nov 2025 · 29.1GPT-5.2GPT-5.2 OpenAI · 11 Dec 2025 · 30.4DeepSeek R1 DeepSeek · 20 Jan 2025 · 11.4QwQ 32B Alibaba · 5 Mar 2025 · 9.5DeepSeek R1 0528 DeepSeek · 28 May 2025 · 13.1ZGLM-4.6 Z.ai · 30 Sep 2025 · 18.5Kimi K2 Thinking Moonshot AI · 6 Nov 2025 · 22.0DeepSeek V3.2 DeepSeek · 1 Dec 2025 · 21.5MiMo-V2-Flash Xiaomi · 16 Dec 2025 · 22.4ZGLM-4.7 Z.ai · 22 Dec 2025 · 22.2
11 new #1 modelstop score +15.2
8 open, close behind
- 2026GPT-5.2GPT-5.2 OpenAI · 11 Dec 2025 · 30.4GPT-5.3 CodexGPT-5.3 Codex OpenAI · 5 Feb 2026 · 32.5GPT-5.4GPT-5.4 OpenAI · 5 Mar 2026 · 39.0Claude Opus 4.7Claude Opus 4.7 Anthropic · 16 Apr 2026 · 40.7Claude Opus 4.8Claude Opus 4.8 Anthropic · 28 May 2026 · 41.8Claude Fable 5Claude Fable 5 Anthropic · 9 Jun 2026 · 49.6Claude Opus 5Claude Opus 5 Anthropic · 24 Jul 2026 · 50.8Claude Fable 5.1Claude Fable 5.1 Anthropic · 1 Sep 2026 · 53.4Claude Opus 5.5Claude Opus 5.5 Anthropic · 22 Sep 2026 · 57.6Kimi K2.5 Moonshot AI · 27 Jan 2026 · 23.5ZGLM-5 Z.ai · 11 Feb 2026 · 27.9ZGLM-5.1 Z.ai · 7 Apr 2026 · 26.1Kimi K2.6 Moonshot AI · 20 Apr 2026 · 27.0MiMo-V2.5-Pro Xiaomi · 22 Apr 2026 · 26.0DeepSeek V4 Pro DeepSeek · 24 Apr 2026 · 30.4MiniMax-M3 MiniMax · 1 Jun 2026 · 29.2ZGLM-5.2 Z.ai · 16 Jun 2026 · 33.7Kimi K3 Moonshot AI · 16 Jul 2026 · 43.6DeepSeek V4 Flash DeepSeek · 31 Jul 2026 · 34.3Qwen3.8 Alibaba · 12 Aug 2026 · 39.9Motif 3 Motif Technologies · 12 Aug 2026 · 33.6DeepSeek V4 Pro 0813 DeepSeek · 13 Aug 2026 · 36.0Qwen3.8 27B Alibaba · 14 Aug 2026 · 33.7ZGLM-5.3 Z.ai · 18 Aug 2026 · 44.8ZGLM-5.3 Flash Z.ai · 26 Aug 2026 · 41.8Qwen3.8 Flash Alibaba · 26 Aug 2026 · 39.8DeepSeek V4.1 Flash DeepSeek · 10 Sep 2026 · 39.5MiMo-V2.6-Pro Xiaomi · 21 Sep 2026 · 46.3
8 new #1 modelstop score +27.2, so far
19 open, close behind
A tool tied to one lab is a bet that the lab stays on top. Charrette routes each step to whichever model is best at it now, open models included, and the project keeps its memory when the lead changes hands.
Artificial Analysis Intelligence Index v4.3.2, read 25 September 2026: releases that beat every earlier model, and open-weights releases that matched the top of six months before. Most scores before 2026 are their estimates on the current index. Top today: Claude Opus 5.5, 57.6.
The market
Everyone is building this layer, inside their own product.
Agents are interchangeable and the best model changes hands every few weeks. What lasts is what a project knows and how its work moves, so every vendor wants to own that. It validates the category, and it is why none of them will make it portable.
| Product | Which agents | Which models | What runs the work | Open source | Where memory lives |
|---|---|---|---|---|---|
| GitHub CopilotMicrosoft, OpenAI’s largest backer | Copilot, Claude and Codex, inside GitHub | 20+ from five labs, GitHub’s pick | Agentic Workflows, in preview: an event starts one agent | The editor extension, the workflow tool | Inside GitHub |
| CursorOwned by SpaceX, with xAI | Cursor’s own | About a dozen, Grok and Composer first | Automations: an event starts a Cursor agent | No | Inside Cursor |
| DevinCognition, owner of Windsurf | Devin only | The widest: 160+, through Cognition | Automations and playbooks, for Devin only | No | Inside Devin |
| Claude Code | Claude Code | Anthropic’s | Routines, in preview: an event starts one run | No | Its own files |
| Codex | Codex | OpenAI’s, or a local open model | Automations: scheduled runs in the app | The CLI | Its own files |
| Charrette | Any, mixed within one task | Any, including self-hosted | Next step decided from evidence, across labs | All of it | With the project, in an open format |
Decided by the vendor Some choice, inside their product Decided by you · simplified, September 2026
Every automation here is a trigger that starts one agent on one job. None decides the next step from what the last one found, across labs, and none keeps what was learned where the next tool can read it.
Why open source
Open source is the base. Everything we sell is built on top.
- 01 · The baseThe whole product, open source, on your desktop.
Project memory, the coordinator and the agent adapters. Free for as long as you like, with your own agents and subscriptions. What the project knows is plain files in an open format, next to your code.
- 02 · Built on topCloud for anyone who would rather not run it.
Small teams and large ones get shared memory, hosted runs and approvals from anywhere. Enterprises get it on their own infrastructure, supported. The same core runs underneath, so nothing is lost moving up or back down.
- 03 · How it growsDevelopers bring it in. Teams pay for it.
Developers adopt it free and pull in the cloud when their team shares a project. Every new adapter and every new user makes it more useful to the next developer.
- 04 · What a lead checksWhere project knowledge lives is a choice for years.
So a lead asks hard questions before saying yes. Open source answers each one with something they can check, not a promise.
- payments-service/
- src/
- .charrette/
- memory/
- token-rotation.md
- refresh-window.md
- money-units.md
- tasks/418/
- graph.json
- evidence/
--- status: settled source: your decision · task 418 agents: claude-code, codex --- A failed token rotation retries once, then fails the request.
- Runs on your desktop
- Plain files, open format
- Your agents, your subscriptions
Organisations · licence
Teams · per seat
Anyone · free
Same open core
How it grows
- 1A developer installs it
- 2Uses it on a real project
- 3The team shares the project
- 4The team moves to Cloud
More adapters and more users bring the next developer in
- ✓Can we read what it does?
All of it: memory, coordinator and every adapter.
source, every line - ✓Where does project knowledge live?
In our repository, as Markdown, reviewed in pull requests like code.
.charrette/memory/*.md - ✓Does our code leave our network?
Only to the model providers we already use. It can run entirely on our own infrastructure.
self-hosted - ✓Whose keys, whose bill?
Our existing subscriptions and API keys. No markup on compute.
bring your own keys - ✓What if the vendor goes away?
Pin a version or fork it. The memory still opens in any editor.
fork · pin · export
No lock-in to weigh
The business
Charge for coordination, never for compute.
ComputePaid to the labs through the subscriptions and keys your team already has. We add nothing to it.
CoordinationPaid to Charrette for what no single agent gives you: one project record, sync, governance and support.
Open source
Free forever
One developer, on their own machine
- The full memory and coordinator
- Any agent, any model
- Local-first, open format
Moves up when a second person joins the project
Charrette Cloud
Per seat monthly
Teams, small and large, on one project record
- Shared memory across people and machines
- Decisions with authors and history
- Hosted runs, approvals from your phone
Moves up when policy, audit or residency apply
Enterprise
Licence annual
Organisations, on their own infrastructure
- SSO, audit and retention
- Knowledge shared across projects
- Support and security review
Expands with every project that adopts it
Adds to an AI budget instead of competing for it, and stays out of a price war with the labs. The model behind GitLab, Sentry, PostHog and Tailscale: free and complete for the individual, paid where a team needs sync, control and someone to call.
Where we are
The thinking is done. This round ships it.
Done
Thesis and research- The core thesis
- The research note and prior art
- Most of the research
Now
Design and experiments- Designing the prototype: board, conversation, knowledge
- Experiments with the coordinator and project memory
- The first working prototype, in progress
With this round
Ship the open-source app- A local desktop app anyone can run
- Project memory with status and source
- Adapters for the major agent tools
- A review and repair loop across labs
With the next round
Cloud and team supervision- Secure containers for long-running agents
- Approvals from your phone
- Memory shared across people and projects
- Self-hosted for enterprises
Where the capital goesNeither release happens without it. This round funds the open-source app; the cloud needs the next.
The project persists. The coordinator understands. Agents come and go.
Research noteThe missing project layer: the full argument, in depth→Charrette · open source
Charrette