Delegate the work. Review the diff.
The MonkeysCode Code Agent runs AI coding agents across your projects without an editor in the way. Start a run, do something else, come back to a reviewed change with tests already green.
Same engine as the IDE. 50+ tools, four execution environments, and every run signed and revertible.
Included in every plan, no extra cost.
Your laptop is not the only place code lives.
Most agent tools assume the code is on the machine you are sitting at. Real work is not like that.
Local
Full capability on your own machine. Semantic search, interactive shell sessions, background processes, checkpoints, OS-level sandboxing.
SSH
Point the agent at a remote development box. It translates every command over SSH and runs there, not here.
Cloud
Execute inside a managed container workspace, sandboxed by default.
Remote Control
Connect over WebSocket to a peer machine running the remote host agent. Useful for a build server, a test rig, or a colleague’s environment.
Each backend declares its real capabilities honestly, and the agent's system prompt and the UI adapt accordingly. Semantic search is not available over SSH, so the agent is told that rather than being allowed to attempt it and fail.
Decide how much rope to give it.
Plan
The agent produces a structured plan before touching anything. Steps with dependencies, grouped into phases, with status tracking as it executes. You review, add notes per step, and approve or reject.
Manual
Ask before every change.
Accept edits
File edits apply automatically, commands still ask. These are separate permissions because most people trust an agent to edit a file long before they trust it to run a shell command.
Auto
Full autonomy inside the sandbox.
Switch with a keystroke. Nothing with a side effect happens outside the mode you chose.
Every file it touches is snapshotted first.
Before the agent's first mutation of any file, the original is captured. If the file did not exist, that is recorded too.
Selective revert. Roll back one file or an entire run.
Revert planning. See exactly what will be restored or deleted before it happens.
Honest about what it cannot undo. A recursive directory delete is flagged as unrevertable rather than silently skipped.
Workspace scoped. Snapshots record their workspace root, so a revert cannot reach into another project.
“We deliberately do not use git stash for this. git stash push -u would destroy your own uncommitted work alongside the agent's. Your changes are yours.”
It checks its own work, and only reports what it broke.
After edits, the agent runs your project's type-checker or linter automatically.
The important part: it captures a baseline at run start and reports only new diagnostics. If your codebase already has 400 warnings, you do not get 400 warnings. You get the ones the agent just introduced.
It detects the right checker per language, scopes to affected packages in a monorepo, and retries once on transient environment failures rather than reporting a false problem.
Baseline-aware diagnostics
Combined with test-gating, a change that breaks something does not reach your working tree.
Research threads that outlive the run.
Some work is not one task. It is a question you chase across days and a dozen runs.
An investigation is a long-lived thread scoped to a project. Runs are linked to it. Findings accumulate. And the output is not locked in an app database — it is real markdown under .monkeyscode/docs/, git-tracked and editable in any editor you like.
Nobody else has a research primitive
above the level of a single conversation.
Everything the agent can reach.
Files
Read with ranges and auto-truncation, create, delete, rename, search-and-replace editing with exact, fuzzy or regex matching, unified diff patches.
Code intelligence
Go to definition, find references, file outline, diagnostics, rename symbol across the codebase, code actions, and vector-based semantic search against a local index.
Search
Ripgrep with full regex and include/exclude patterns, glob discovery, directory listing.
Execution
Shell commands with timeout and CWD, test running with framework auto-detection, interactive session input and screen reading, background process management.
Git
Status, diff, log, blame, show, commit, branch creation, stash and stash pop.
Web
Search, URL to markdown, raw fetch.
Orchestration
Spawn parallel subagents, ask the user mid-run, timers, structured todo lists, persistent memory across sessions, and search over your own past conversations.
Extensibility
MCP servers from .monkeyscode/mcp.json, plus custom tools defined per workspace, all routed through the same permission policy as built-in tools.
Every tool is permission-gated. Nothing with a side effect runs without either your explicit approval or a mode you deliberately enabled. Nothing runs outside the sandbox.
It runs inside a box, on every platform.
| Platform | Mechanism |
|---|---|
| macOS | Seatbelt |
| Linux | bubblewrap |
| Windows | AppContainer |
In workspace_write mode, writes are constrained to the workspace and temp, and network access is denied unless approved.
Secret scanning. Pre-commit detection blocks findings, and secrets are redacted from tool output.
Untrusted content handling. Content pulled from the web or other untrusted sources is fenced and classified, with prompt-injection detection and redaction before it reaches the model.
Policy engine. Per-tool permission tiers evaluated as allow, deny or ask, configurable at workspace and organisation level.
Sandboxed by default
Seatbelt · bubblewrap · AppContainer
Built to stay coherent past 100 messages.
Token estimation per message, compaction when the window fills, and a live meter showing usage against the model's capacity.
Messages are classified by stability into immutable, session-stable and per-turn volatile tiers, and cache breakpoints are placed to maximise provider cache hits. That is why about 90% of context is served from cache and why frontier requests average around $0.06.
Workspace rules from .monkeyscode/ are loaded into every run, so project conventions apply without being restated.
context served from cache
avg per frontier request
Watch many agents without losing the thread.
Left rail for runs, investigations, docs and attachments.
Right context sidebar with six views: plan, gaps, subagents, artifacts, tasks and terminals.
Streaming diff viewer shows edits as they are generated. Batch edit review lets you accept or reject per file.
Real-time cost tracking in the footer.
Command palette on \u2318K.
Everything persists locally in SQLite. Conversations, messages, events, checkpoints and investigations survive restarts.
Built for supervision
Watch many agents without losing the thread
Included in every plan. Free to try for 30 days.
The Code Agent, the IDE, and unlimited Capuchin. No credit card.