requests

• agent memory1, shipped
• agent tied to a repo1, shipped
• byte-faithful round trip1, shipped
• everything behind an info button1, shipped
• idk1, not yet
• live model picker1, shipped
• per-user auth1, shipped
• per-version telemetry1, shipped
• publish on the timeline1, shipped
• reasoning in chat1, shipped
• a search key0, not yet
• a workspace to read0, not yet
• ask as a waitpoint0, not yet
• billing and top-ups0, shipped
• bring your own key0, shipped
• chains out of the editor0, shipped
• comments0, shipped
• compiled skills0, not yet
• enforced permissions0, not yet
• eval execution0, not yet
• evals and judges0, shipped
• four mascots0, shipped
• mcp tools at runtime0, not yet
• per-user mcp tokens0, not yet
• queue visibility0, not yet
• run history0, not yet
• sharing0, not yet
• synced notes as memory0, shipped
• the cute loader0, shipped

shipped not yet

1. Junper-user auth29 Jul
2. Junagent tied to a repo29 Jul
3. Junbyte-faithful round trip29 Jul
4. Junreasoning in chat29 Jul
5. Juneverything behind an info button29 Jul
6. Junlive model picker29 Jul
7. Junpublish on the timeline29 Jul
8. Junper-version telemetry29 Jul
9. Junagent memory30 Jul
10. Junidk31 Jul

1. A comparison is kept and reopened at /compare?run=, with a recent listapp10 Aug20 Aug
2. Pi, Claude Code and Codex carry their own marks in the configapp10 Aug20 Aug
3. A run uses the connection the user actually connected underruntime20 Aug20 Aug
4. A sandbox reaches MCP with a signed ticket, never a credentialruntime19 Aug20 Aug
5. MCP tools reach a run, single-tenant only until per-user tokens landruntime29 Jul20 Aug
6. An agent can actually call an MCP tool: proven end to end against Slackruntime19 Aug20 Aug
7. An MCP run gets a session scoped to the toolkits its config namesruntime19 Aug20 Aug
8. A token chip shows the model maker's markapp10 Aug20 Aug
9. A model the key cannot use is explained, with what to changeruntime10 Aug20 Aug
10. Compare runs each variant once; sampling waits for evalsapp10 Aug20 Aug
11. A run's MCP access names one user and carries only their connectionsruntime29 Jul20 Aug
12. A follow-up prompt stays on screen instead of becoming the word Thinkingapp10 Aug20 Aug
13. The publish dialog's path warning reads as a sentence againapp10 Aug20 Aug
14. A version suggests its own name and change list, free, before you save itapp10 Aug20 Aug
15. A turn shows its tool calls where they happened, not stacked above the replyapp10 Aug20 Aug
16. Draft, live and retired are named and explained in the versions tableapp10 Aug20 Aug
17. Web search works under Claude Code and Codex, not only Piruntime10 Aug20 Aug
18. A new agent starts with every native tool enabledapp10 Aug10 Aug
19. A comparison opens as a playground: both conversations live, one message to allapp10 Aug10 Aug
20. A comparison carries only the documents it can reach, and says why a run failedapp10 Aug10 Aug
21. Connectors is one list: popular, three filters, and state carrying the actionapp10 Aug10 Aug
22. One ranked chart of total cost by version, versions with no spend left outapp10 Aug10 Aug
23. Cost by token class is a ring showing each class's share of the billapp10 Aug10 Aug
24. Every feature is one line on the requests page, with the day it was asked forlanding9 Aug10 Aug
25. An MCP app is one switch with its logo, and takes all its toolsapp10 Aug10 Aug
26. The model dropdown shows the maker's logo and sorts newest firstapp9 Aug10 Aug
27. A picked model is stored in the spelling Pi accepts, not OpenRouter'sapp10 Aug10 Aug
28. Every model maker's logo is served from our own origin, 42 of themapp9 Aug10 Aug
29. A skill downloads instantly as the SKILL.md it came from, and the list searchesapp10 Aug10 Aug
30. 104 skills from the machine's own folders are in the appapp10 Aug10 Aug
31. Thinking levels are the ones a model really serves, across all seven Pi hasapp10 Aug10 Aug
32. Tokens by model is a coloured bar chart that draws itself on arrivalapp10 Aug10 Aug
33. Five use cases named on the site: administration, construction, health, finance and ecommercelanding9 Aug10 Aug
34. Secrets and the model picker draw every vendor logo from one generated mapapp10 Aug10 Aug
35. web_search runs on a real Exa key instead of a free tier that 429sruntime6 Aug10 Aug
36. A runtime can be drafted and entered as a variant before it is promotedapp8 Aug
37. A skill can be published globally rather than only to the account that wrote itapp8 Aug
38. An image can be attached to a chat messageapp8 Aug
39. A conversation can be named, and one compaction default covers every agentapp8 Aug
40. Compaction and retry are a real .pi/settings.json, behind an Advanced sectionapp7 Aug
41. A key can be brought for every provider Pi documents one for, with its markapp7 Aug
42. Input, output, cache read and cache write are metered and shown as four ratesapp7 Aug
43. A Q&A page, one paragraph and three sentences per answerlanding7 Aug
44. A download is a runnable folder for the harness the agent declaresapp4 Aug
45. A publish never overwrites a file it did not writeapp4 Aug
46. The agent form picks its harness first, and nothing is deleted on a switchapp3 Aug
47. A published config runs under local pi, from a compiled artifact and a READMEapp3 Aug
48. A version can be named, and billing shows two bars instead of one figureapp3 Aug
49. A turn runs the compiled agent file, so every declared field reaches the runruntime2 Aug
50. True vendor cost is recorded per run and per message, beside what was chargedapp2 Aug
51. A declared permission policy is enforced, with .env reads denied as a floorruntime2 Aug
52. A skill is a document you can write, and publishing carries it into the runapp2 Aug
53. A chat turn is 8.4s faster, from an image that pays Pi's first load onceruntime2 Aug
54. Your portrait is an egg in a nest, and the nest is twig, green or goldapp1 Aug
55. One shell for every place, so connections and billing keep the railapp1 Aug
56. Version history is the diff, and the modes moved into the railapp29 Jul1 Aug
57. Chains are authored in the editor again, and run from itapp31 Jul
58. Publishing an agent with no repo offers to create one, private by defaultapp31 Jul
59. Every account picks a handle, and an agent lives at /@handle/nameapp31 Jul
60. Memory is its own page, with a pre-upload review and a secret scannerapp31 Jul
61. Notes sync both ways, with a three-way reconcile and conflicts as siblingsapp31 Jul
62. One Publish that pins a version and commits it, with a Chat tab per versionapp29 Jul31 Jul
63. Requests are listed oldest first with the first name and picture of the senderlanding31 Jul
64. An agent's memory file is seeded before a turn and stored after itruntime30 Jul30 Jul
65. Ask for something from the account menu, one vote per account per topicapp30 Jul
66. Top-ups through Stripe Checkout, auto-recharge, seats and a ledgerapp30 Jul
67. Pricing, cabinet, demo, status and requests, each a page off the home pagelanding30 Jul
68. Egg, hatching, chick and hen, built in Three.js and placed where each workslanding30 Jul
69. Every chat turn and chain run leaves as an OTLP trace, subagent spend nestedruntime30 Jul
70. The model list is the live OpenRouter catalogue rather than a hardcoded oneapp29 Jul30 Jul
71. A model's reasoning is shown in the chat as its own blockapp29 Jul30 Jul
72. The requests cloud counts how many different accounts asked for a thinglanding30 Jul
73. An agent is bound to a repo, and Publish writes a real commit to itapp29 Jul29 Jul
74. Form, YAML and JSON are one object, and a round trip changes no bytesapp29 Jul29 Jul
75. Anchored comment threads on a config, with agents as first-class authorsapp29 Jul
76. GitHub and Google sign-in, with every agent a row only you can readapp29 Jul29 Jul
77. Connect and disconnect GitHub and Composio per account, secrets in the vaultapp29 Jul

1. web_search has a real key instead of a free tier that 429sruntimenot yet
2. A chat sandbox holds more than memory, so an agent has files to readruntimenot yet
3. An ask pauses the run and waits for an answer instead of failingruntimenot yet
4. An eval definition actually scores somethingappnot yet
5. You can see the queue and how deep it isappnot yet
6. Past runs are listed and can be reopenedappnot yet
7. An agent can be shared with somebody elseappnot yet
8. A turn can be stopped once it has startedappnot yet