What's new
in each release.

Conversations are saved. Each conversation is written to your computer as it happens, so closing Lamied, a crash or a power cut doesn't lose it, and opening a folder again brings back the conversation you had open there. /resume lists the folder's other conversations, newest first. /fork continues a copy of the conversation, whole or up to one of your messages, so you can try another approach and keep the first.

Hooks. A project can run its own commands while the agent works: before it uses a tool, to block an action such as committing a secret; after it uses one, to lint every edit, say; and when it says the task is done, to send it back to work while tests still fail. Lamied asks before running a project's hooks for the first time, and again when they change. Hooks use Claude Code's format, and a project without .lamied/hooks.json uses those in .claude/settings.json, so hooks a team already has carry over. See Hooks.

Rules for each folder. A folder can have rules of its own, in a LAMIED.md, .lamied/rules.md or AGENTS.md inside it, so each package of a monorepo can say how it is built and tested. The agent reads them the first time it works in that folder. A project set up for Claude Code works as it is: where none of these files exists, Lamied reads CLAUDE.md.

Also in this release:

• /init writes an AGENTS.md that says how to build, test and work in the project. /review [base] reviews the changes since HEAD or a branch for bugs, risks and slower code.
• Long conversations keep working. Lamied summarizes the start of a conversation once and reuses that summary, so a request stays the same size however long the conversation runs. Before, a very long conversation stopped accepting messages.
• A dropped connection no longer ends a task. The agent reconnects, up to four times over about 40 seconds, and sends the step again with the work already done kept.
• On the Team plan, agent tasks no longer fail when hosted requests are rate limited. They wait their turn, and the status line says for how long.
• Replies start sooner after a pause. Connections stay open for a minute: measured from Ethiopia, a request after 5 to 60 idle seconds reaches Lamied's server in about 0.3 s instead of 0.8 to 0.95 s.
• Agent tasks cost less, because more of each request is served from the provider's prompt cache. In simulated tasks of 25 to 80 steps, input cost fell by 35–62% on Claude and 15–40% on GPT-5-class models.
• Long file writes show their progress, such as "Writing src/app.ts (120 lines so far)…", instead of "Thinking…". MCP servers start while you type, and a task waits at most 5 seconds for one still starting, instead of 30.
• Closing a file never loses unsaved changes. With autosave on, the file is saved first; otherwise Lamied asks whether to save, discard or cancel.
• A model that wrongly says it has no tools is sent back to work, and its refusal is taken out of the chat. After three refusals in a row, the task stops and says why.
• With a Voyage key, semantic search uses Voyage's voyage-4-lite model. Each workspace is indexed again once, the first time this version opens it.
• A paid account no longer reads as Free when two status checks overlap.

Lamied Pro is free until 28 October. Sign in and the AI features are unlocked on the Developer plan, with your own API keys or a local model. Nothing is on sale until then, including the Team plan.

Paying by bank or Telebirr. After the free period, choosing a plan opens a checkout with CBE, BoA and Telebirr. Transfer the plan's ETB price, upload a screenshot or statement, and Pro unlocks as soon as the payment is approved, usually within about 30 minutes. Update to this version to pay: earlier versions can't open the new checkout.

Also in this release:

• Plan cards, the Team total and the Student note show ETB prices only.
• The account menu no longer has an operations section.

Three Lamied Pro plans. Choose the one that fits in the AI panel, or from Upgrade in the status bar:

| Plan | Price a month | | | --- | --- | --- | | Student | ETB 499 or 5 USDT | For current students, once your student ID is approved | | Developer | ETB 1,199 or 12 USDT | For one developer, with your own API keys or a local model | | Team | ETB 1,799 or 18 USDT per member, 3 or more | Adds OpenAI, Claude and Gemini models hosted by Lamied, with no API keys |

• Student: choose a photo or scan of your student ID, check it, and submit it for review. You can replace it while it waits. We email you the result, and once it's approved the panel offers the Student plan.
• Team: the person who pays adds members by email under AI settings → Team. Members choose Lamied (Team plan) as their provider and use the hosted models with no key: Lamied keeps the provider keys on its own server. Each team gets ETB 2,000 of hosted-model use for every 3 members a month, every request is metered, and the panel shows how many ETB are left after each response.
• The waitlist: if you can't pay right now, join the waitlist below the plans. When your turn comes you get a free month of Pro, an email saying so, and the panel unlocks on its own.

Also in this release:

• The app connects to Lamied's new server, which fixes email verification links that reported "This link has expired" right after sign-up.
• Your own Gemini key is now sent in a request header rather than in the address.

An agent that runs your project, not someone else's. The AI agent now finds how a workspace runs by searching it (Compose files, Dockerfiles, Procfiles, package scripts, Python entry points, deployment configs) instead of reading the whole codebase, and reports what is already running on each port and which process owns it. It refuses commands that reach into a different project's folders, asks before stopping processes, and no longer reruns a command that failed when nothing has changed since.

Also in this release:

• OpenRouter is a new provider, including Apodex 1.1 Mini. What each model supports (tools, images, context size) is read from the provider's catalog and learned from its errors, so every model gets every tool.
• Complexity analysis is 2.4x faster. Files are analysed on several workers at once: 18.1 s down to 7.5 s on an 871-file project, with identical results.
• Long commands such as image pulls and builds keep running while they print progress, instead of being cut off at a fixed deadline.
• Web addresses in the chat, such as a server's http://localhost:8000, open in your browser.
• The chat panel shows paths relative to the project and no longer repeats the same output; a repeat collapses to one line.
• Plan Mode only asks for approval once there is a plan to approve.
• A greeting no longer costs a second model call, and the call count and tokens shown now include every call the task made.

Network analysis. The complexity dashboard's Network tab now estimates what a file's network and database calls cost as input grows: how many requests each function makes per call (O(1), O(n), O(n^2)), whether they run one after another or all at once, and how many rows its reads return. It reads real syntax trees, follows calls across files through imports, types and callbacks, and reports problems such as N+1 queries with the line, an explanation and a fix. Nothing is executed. Findings can be explained by the AI or added to the AI chat.

Also in this release:

• /fim chooses the provider and model for suggestions while typing, separately from the chat model.
• The AI agent keeps a trace of the context it read and searches projects more precisely.
• New application icons on Linux and Windows.

Lamied now installs on Windows. This release adds a Windows x64 installer alongside the Debian package. The installer adds lamied to your PATH, and the uninstaller removes it. Windows builds link the native engine against static vcpkg libraries, so no Visual C++ redistributable is needed.

Also in this release:

• Inline code completion backed by the language server, and a fix for suggestions that could linger after the cursor moved on.
• Video upload, with its own page on the website.
• Mistral joins the configurable AI providers.
• The download page offers every installer the release carries, so Windows and Linux visitors are each shown their own build first.

Pro unlocks reliably after payment. If the connection dropped or timed out at the moment a payment was approved, the AI panel could stay locked until Lamied was restarted. Lamied now keeps checking until the backend answers, then unlocks. A temporary server error is no longer mistaken for a free account.

This release makes the AI assistant faster and safer to work with, gives the editor its own language intelligence, and adds local payment options and email verification.

Pay with CBE, BoA or Telebirr. Lamied Pro now costs 500 ETB a month by bank or Telebirr transfer, or 5 USDT. Transfer, upload a screenshot or statement, and Pro unlocks as soon as the payment is confirmed, usually within about 30 minutes. If a USDT payment isn't detected, you can attach proof on the same page.

• Pro unlocks in the app the moment the payment is approved; no restart or waiting.

Verify your email. New accounts confirm their email before signing in. The link opens a page on lamied.com.et, and Lamied moves to sign-in by itself once it's opened. If the email doesn't arrive, Resend tells you exactly what happened.

Lamied is available in Ethiopia. The app, the website and payments work from Ethiopian networks.

Inline AI suggestions. While you type, the code around the cursor goes to your selected AI provider, and a suggestion appears in grey. Tab accepts it, Ctrl/Cmd+→ accepts one word, and Esc dismisses it.

• Wrong Python imports are corrected instantly from the modules the file already imports, with no model call. The fix is shown under the line, as in VS Code.
• Suggestions stay small and fast: they use a separate backend route with its own limits and capped output, so typing never uses up the assistant's budget.
• Typing in credential files (.env, private keys, .npmrc, …) never sends anything.
• You can turn suggestions off in AI settings → Editor.

Checkpoints. Before an AI task first changes your workspace, Lamied saves a checkpoint of every file that isn't ignored. Click ↩ on your message to put every file back the way it was before that task. You can also type /undo for the latest task, or /checkpoints to pick any checkpoint.

• Lamied shows which files will change before restoring, and saves the current state first, so a restore can itself be undone.
• Checkpoints live in a private store in the app's data folder. Your project's own git history is never touched.

Language support from the new Extensions view. A new Extensions view in the activity bar installs language servers for live errors, hover and go to definition:

• Python (Pyright)
• TypeScript and JavaScript
• Rust (rust-analyzer)
• C and C++ (clangd)
• Go (gopls)

Details:

• Nothing is bundled or preinstalled. When you open a supported file, Lamied offers to install the matching extension.
• Downloads come from pinned versions, are checked against published checksums, and are installed atomically.
• Every extension can be disabled, re-enabled or uninstalled.
• Language servers run only in workspaces you turn them on for, because they execute the project's own tooling.
• The assistant uses them too: get_diagnostics checks its own edits, and find_references finds usages.

The assistant works more like Cursor:

• Reads and searches fold together. Each run of reads and searches, and the thinking between them, folds into one group: *Explored 6 files, 2 searches ⌄*. Click it to see each step.
• Live summary. A line shows progress with line counts, e.g. *Editing 2 files, explored 6 files +30 −16*.
• Files Changed card. After the answer, each changed file is listed with its icon and +/− counts. Click a file to open it, or use Review to see every diff.
• Answer actions. Each answer has a Copy button and shows its time.
• Pinned question. Your question stays pinned at the top while the answer scrolls.
• Plain-text thinking. Thinking is shown as plain prose, without **, # or other Markdown markers.
• File links. File names mentioned in the chat link to the file, at the mentioned line.
• Changes bar. A Changes +A −D bar above the message box shows what the assistant changed and opens Source Control, where you commit. Switch or create branches from the branch name at the bottom left.

Faster and more capable assistant:

• Precise edits. The new edit_file tool replaces exact text instead of rewriting whole files.
• Open file included. The file open in the editor is included, so "this file" needs no lookup.
• Sub-agents. Research sub-agents investigate in parallel. Worktree sub-agents implement changes in isolation on their own git branches.
• Project rules. Rules in LAMIED.md, .lamied/rules.md or AGENTS.md apply to every request in that project.
• MCP servers. Connect Model Context Protocol servers (issue trackers, databases, docs, browsers) from AI settings → MCP servers. Every MCP tool call asks first unless the server lists it under autoApprove.

Complexity dashboard and git:

• The complexity dashboard is redesigned for the sidebar: file cards sorted worst first, readable details, and the AI's thinking shown live.
• Add to AI chat puts a file's analysis in the chat's message box, ready for your request.
• Run analysis re-analyzes only the files that changed and keeps earlier AI explanations.
• The current git branch is shown at the bottom left. Click it to switch or create a branch; if your changes block a switch, Lamied offers to stash them.

Fixes:

• Temporary AI provider errors (busy, rate-limited) are retried automatically, and errors are shown as a short sentence instead of the provider's raw response.
• The editor no longer throws Missing requestHandler or method: getCodeFixesAtPosition. TypeScript, JSON, CSS and HTML each use their own editor worker again.
• Suggestions can no longer show raw model output or protocol markers.

Full usage, configuration, safety model and limits are in AI features.

This release focuses on making Lamied faster, more stable, and easier to use in daily coding sessions.

Workspace search, symbol graph traversal, AI context retrieval, and indexing now use the native parser and background analysis path more consistently. Large workspaces should feel more responsive, with lower main-thread blocking during analysis and better reuse of indexed results after small edits.

Payment handling is more predictable. Crypto checkout keeps polling through overpaid and confirming states, avoids misleading underpaid messaging once another network is confirming, and limits invoice polling so abandoned historical invoices do not crowd out active payments. The admin payment-sweep panel was verified for manual sweep, confirmation, submitted, retry, and reconcile states.

The desktop editor also adds an Ubuntu Terminal color theme, fixes missing TextMate grammar errors for C++ embedded macros and SassDoc, and restores app zoom shortcuts. Ctrl++, Ctrl+=, Ctrl+-, and Ctrl+0 now work across the editor, terminal, chat panel, and other focused controls.

Unlock with USDT now opens checkout immediately, with a loading indicator while the payment page arrives. Solana and TRON QR codes display correctly in the desktop app. Back cancels the loading view, and late responses cannot reopen it.

The checkout framing fix is limited to the configured backend inside a verified Lamied window. Browser framing protections and the payment page's script policy remain intact. Both network tabs, Copy, and payment-status polling were verified.

• Screenshots for the agent. The AI assistant can open a page in a hidden, isolated browser and look at it, so it checks layout, spacing and colors instead of only reading CSS. Screenshots appear as thumbnails in the chat. Local addresses (localhost, 127.0.0.1) open without a prompt; any other URL asks you first. The browser uses none of your cookies or logins, and Stop cancels a capture in progress.
• Better UI work. The assistant reuses your project's existing design tokens, builds layouts that work from phone width up, covers accessibility basics, and verifies visual changes with a screenshot before calling them done.
• Preview tab. A live view of your running dev server beside Problems, Output and Terminal, with reload and "open in your browser".
• Works with every model. Screenshots are sent to vision-capable models; if a model does not accept images, the assistant continues without them. Only the two most recent screenshots stay in context to keep requests small.

The AI assistant can now answer with real charts, switch models from the chat, and shows its work more clearly.

• Charts in chat. Visual answers render as interactive, themed 2D and 3D charts (lines, bars, heatmaps, contours, surfaces, and paths on them) instead of ASCII art. Curves and surfaces are computed from formulas with a safe math parser; nothing the model writes is evaluated as code.
• Tables. Markdown tables render with alignment, formatting, and math.
• Slash commands. /model picks a provider, then a model from that provider's live model list; /model <id> sets one directly. /clear, /stop, /settings, and /help are available from a command menu.
• Clearer agent activity. Thinking streams smoothly and folds into "Thought for Ns". Terminal output is a collapsible log. Plans appear as a live checklist. Rows no longer overlap or clip when the transcript is long.
• Loading indicator. The animated Lamied mark replaces spinning loaders.
• Fixes. Gemini no longer rejects continued agent runs with "Requests ending with a model turn are not supported", and Claude no longer treats them as prefill. The "Allow commands without asking" setting updates as soon as a workspace opens.
• Promo codes. Administrators can issue, list, and revoke one-time codes that grant a month of Pro access; anyone signed in can redeem one.

A more welcoming workspace, from the first launch to the next project.

• New start screen with working project, new-file, and command-palette actions.
• Redesigned sign-in and sign-up with a branded code illustration, clear errors, keyboard focus handling, and protection against duplicate/stale UI requests.
• Responsive dark/light layouts and reduced-motion support, without additional UI frameworks, remote fonts, or animation timers.
• Refreshed AI access and account panels. Local editing remains available without an account; AI access requires the applicable account and subscription.
• Mandatory C++ engine included and validated with all 14 language grammars.
• Generated test reports and environment files excluded from the release archive.