AI that does real work on your machine
with fewer tokens, on your terms
Chat with AI or let an agent edit real files, only in folders you approve. Every model and agent you already pay for, in one app.
- 31–60%
- fewer input tokens in our benchmark
- 13
- model providers in one app
- 0
- servers or telemetry
−60%
input tokens vs OpenCode, 7 tasks on one model
Ask anything, no file access
Summaries, drafts, research, planning. Chat mode never reads or edits your files — it only answers.
e.g.“Sum up these meeting notes in 3 points and draft an email to the team”
Save tokens
Tools load on demand, stale output is shortened, JSON is compacted: 31–60% fewer input tokens for the same work in our benchmark.
You approve every action
Anything with real effects shows a card first, with the full command and real paths to read.
Undo any file
The app remembers your folder before every agent turn. Don't like the result? Undo.
Works with every model and agent you already pay for
Let AI do real work,
safely and for less
Cowork on real files
Let an agent sort files, read documents and write reports, only in folders you allow, read-only or read & write, with every line's diff to review before you accept.
- Folder access
- Permission card
- Checkpoint
- Files changed
Token Saver
Every agent step re-sends the whole pile to the model. Mali makes the pile smaller: measured on 7 tasks, 60% fewer input tokens than OpenCode.
- Tools on demand
- Stale output
- Compact JSON
- Prompt caching
Connectors (MCP)
Find connectors in the official MCP Registry, install them from chat, or sign in with OAuth as Mali Cowork. Keys live in your Keychain, never in chat or the model.
- Notion
- GitHub
- Linear
- Figma
Code mode
Chat and code on one screen. Watch the files the agent edits in real time, run build and tests, fix errors with AI in one click, then get a drafted commit to confirm.
- Explorer + Editor
- Run / Check
- Fix with AI
- Git commit
Code map
Open a code folder and Mali builds a knowledge graph of it in the background. Every agent asks the graph (graph_query, graph_path, graph_explain) instead of opening files one by one, so it lands on the right place sooner. The graph stays in Mali's own folder, never in your project.
- graph_query
- graph_path
- graph_explain
- Never writes to your project
Same work, fewer tokens sent
Every step an agent takes re-sends four things to the model: the system prompt, the built-in tools, the connectors' tools and the history so far. Mali makes each pile smaller, and the work still passes the same checks.
input tokens per task vs Mali's previous agent
vs OpenCode on the same model and tasks
quality checks passed on runs without provider errors
Scroll through each lever and watch a pile shrink
- 0
Before: everything, every step
An agent works in rounds: think, call a tool, read the result, think again. Each round sends the whole pile to the model. Connectors like Canva or Notion bring dozens of tools with long schemas, sent again every step even when the task never touches them.
The baseline we compare against - 1
Connector tools load on demand
Instead of every tool's full schema, the model gets a one-line-per-tool catalog through load_tools and loads the full schema only for what it uses. The tool list only ever grows at its end, so the provider's prompt cache stays warm.
19,185 → 3,242 characters per step · −83% - 2
Stale output is shortened
A page fetched or a file read is sent again with every later step. Output over 2,000 characters and older than 6 messages is shortened in one batch once more than 12,000 characters pile up, not every step. The agent can always run the tool again for the full text.
14 days of real use: re-sent output came 36% from web fetches, 24% from file reads - 3
Compact JSON and short image handles
Pretty-printed connector results carry a lot of whitespace; they're squeezed to one line with the same data. Signed image links of ~830 characters reach the model as 45-character handles while the app keeps the image.
JSON −26% · image links −95% - 4
Prompt caching
The start of each request that repeats the last one (system prompt, tools, history) is cached at Anthropic and billed at roughly 10%. The token count stays the same; the bill doesn't.
Claude models only · not measured in the benchmark - 5
Near the limit? Summarise and carry on
At 75% of the context window the app first shortens old output, then, if that isn't enough, has the model summarise everything before the prompt (goals, decisions, files, what's still open) and continues from the summary.
Kicks in at 75% of the window · not part of this simulation
7 tasks × 2 rounds on the same model
gpt-oss:120b on Ollama Cloud · same sample folder · a mock connector with 32 tools · scored by plain checks (right numbers, files present, tests pass), not by another model.
average input tokens per task · gpt-oss:120b
Per task
The full report, including the parts that don't flatter us
| Engine | Total input | Billed at full price | Quality checks | Provider errors |
|---|---|---|---|---|
| OpenCode | 72,314 | 8,330 | 94% | 0 |
| Mali (before) | 42,025 | 4,057 | 98% | 1 |
| Mali + Token Saver | 28,849 | 5,228 | 93% | 2 |
Total input counts every token sent, cached or not. Providers bill cached input at about a tenth, so "billed at full price" is closer to your invoice: there Mali + Token Saver is 37% below OpenCode, but higher than Mali before (5,228 vs 4,057), because shortening old output changes the cached prefix. Quality is the share of plain checks passed over 14 runs; the 93% includes 2 runs cut short by provider 500 errors, and the lowest task (CRM data to files, 50%) is one of them.
In 2 of 14 Token Saver runs and 1 of the previous agent's runs, the model provider returned a 500 error mid-task. Runs without errors passed 100% of checks either way · Set MALI_TOKEN_SAVER=0 to compare yourself.
Promises you can check in the source
API keys live in your OS keychain, not a text file, and the model never sees them.
No analytics, no telemetry, no account and no Mali server.
Anything with a real effect waits for your click first.
Chat mode can't touch files at all, and a chat stays in one mode for good.
Before every agent turn the app snapshots your folder so you can undo.
The AI never installs a connector itself. It suggests a card; you review and confirm.
Connect everything you already use
OpenCode, Codex, Gemini CLI, Cursor Agent, the APIs of 13 providers, and hundreds of connectors from the official MCP Registry.
Free, runs on your machine, and your data stays yours.
Download and run—no account required
Mali Cowork is free. You pay model providers directly, or use OpenCode / local Ollama free tiers.
Current builds are not code signed / notarized
macOS may block the app as an unknown developer → right-click the app → Open → Open
Windows SmartScreen may warn → More info → Run anyway
Early-access limitation, not a sign the file is malicious. Developer signing comes in a later release—or build from source if you prefer.
After first launch
- 1
Install OpenCode
Free models included—no signup—or skip to step 2 if you already have keys.
- 2
Add your provider API keys
Stored in the OS keychain / credential manager, not plain text files.
- 3
Enable MCP tools you want (optional)
Install from chat. Toggle each server anytime.
The app walks you through step by step.
Questions worth asking—and honest answers
What the app can and cannot do, without overselling safety.
Keys live in the OS credential store
macOS Keychain · Windows Credential Manager—not plain text files. Models never see keys; they stay in HTTP headers, not chat context.
No telemetry
No analytics, crash reporter, accounts, or Mali server—verify in source.
Rust 141 tests · TS 28 tests passing
Coverage for permission decisions—policy, OAuth, path resolution.
Question by question
Usable every day, with limits you should know
This table is not marketing—it is here so you can decide clearly.
| Topic | Status |
|---|---|
| Core features (Chat, Cowork, MCP, Git, Checkpoint) | Available |
| Key storage / permissions / OAuth | Test coverage (Rust 141 · TS 28 passing) |
| Code signing & notarization | Not yet—first open via right-click → Open |
| Linux credential store | Owner-only file instead |
| Third-party security audit | Not done yet |
Good fit for
Developers and power users who want one hub for every model/agent and clear control over what AI can touch.
Not yet for
Organizations that require an externally audited app, and jobs that must run 24/7 in the cloud: Mali Cowork only works while your computer is on.
Small things people ask
What people ask about AI agents on their own machine
How a desktop AI agent actually works, how it compares with Claude Cowork, and where the token savings actually come from
Let an agent do the work
No account, no telemetry. Download, open, and pick the first folder you want help with.