Settings
/config opens it. Eight tabs:
Status, General,
Project, Display,
Advanced, Models,
Usage and Stats.
Tab switches, ←
→ changes a value, Enter
opens what a line offers (a text field, a list, a page of its own).
Most values are global and apply to every project and session. The Project tab holds the few that belong to one project and override the global one. The Models tab can additionally hold a choice for one session only. Where a value can be set from an environment variable, that is listed at the end of this page — see Environment variables.
Status
What is running right now, one line each and read-only: Project, Session, Working dir, Provider, Model, Effort, Cache TTL, PUM (on or off), PUM harvest, PUM sync (current, or “local is behind”), Permissions (the mode), Licence (state, tier, the date updates are included until) and Machine. From here l opens the licence page (see Licence & updates) and h the health check, which lists what Nemocode could verify about your installation (✓ fine, ○ optional or unknown, ✗ needs attention).
General
The settings you touch in normal use.
| Setting | What it does | Default |
|---|---|---|
| RTK token saver | Compresses long shell output before the model sees it. Needs the helper on your PATH; the installer brings it. | on |
| Re-read dedup | When the agent reads a file it already read and nothing changed, it gets a short note instead of the whole file again. | on |
| Predictive preload | Experimental: load the files you usually need before the agent asks. Saves steps when your habits are stable, adds noise when they are not. | off |
| remember tool | Whether the agent can note something down mid-mission as its own step. Off means knowledge is only noted as text inside the answer, which costs no extra request. | off |
| Verify subagent | A second agent checks the finished
work before a mission may end; a failed check sends the mission back for
another round, up to twelve times, after which the mission ends with the
failing verdict. Very small changes (fewer than three writes) skip the
check. Off by default. A mission started with
/goal gets the independent check regardless of this switch.
If the environment variable NEMOCODE_INDEPENDENT_VERIFIER is set
(0 means off, anything else on), it decides: the line then
reads “set by …” and cannot be changed here until you unset it and
restart. | off |
| Auto-approve plans | Build straight on without waiting for your approval. For unattended runs; the plan is still written down in the project. | off |
| After a plan is approved | Which mode it continues in: auto mode asks nothing, or accept edits but still ask before shell commands. | auto mode |
| cross-session talk | Whether sessions may send each other messages. Off means the ability is not offered to the agent at all. | off |
| PUM loading | The project memory (PUM) is either an overview plus lookups on demand (map) or everything upfront (full). | map |
| PUM harvest | Who gathers new facts: the agent's own notes (agent), the curator that harvests the session (curator), both, or nobody (off). | agent |
| Harvest trigger | Whether the curator harvests every mission by
itself (automatic) or only on /harvest
(manual; missions wait in the harvest session until then). See
the harvest session. | automatic |
| Harvested by | Per project: who harvests the sessions of Claude Code, Codex and opencode — Nemocode's curator, or the agent itself. See Other coding agents. | Nemocode |
| Session access | Which past sessions the curator may read, from which projects. Enter opens the list; it is also reachable with e in the project and session lists. Per project. See the curator. | nothing granted |
| External agents | What Claude Code, Codex, opencode and Claude Desktop may do in each project: off, read, read + propose or curator. Also x in the project and session lists. See Other coding agents. | off |
| Integrations › | Opens the window
Integrations (nemocode integrate): Claude Code, Codex, opencode and
Claude Desktop, each shown as connected or not, with the path of the
configuration file Nemocode would write or has written. i
connects the marked one and u disconnects it; both
ask you to press the key a second time. This only changes that
tool's configuration — which projects it may use is set under External
agents. Needs the nemocode command on your PATH. See
Other coding agents. | — |
| PUM sync | Share this project's memory with colleagues: remote, push and pull, auto-push interval (seconds; 300 unless you change it, 0 is off), the host name colleagues see first in invites, review rules, reviewers, required checks and check commands. The source of a team also sets the gateway, the security policy and what members receive there. See Sharing with a team. | — |
| Embedding model | The model that finds related facts. Enter lists them with size and whether they are downloaded; a download runs in the background. | qwen3-embedding:0.6b |
| Embedding auto-pull | Whether Nemocode downloads a missing memory model by itself. | on |
| Embedding startup refresh | Whether the installed memory models are checked every time the server starts. | on |
| Compaction ask | Whether Nemocode asks before summarising a long conversation: off, always, or smart, which asks only when the summary itself is uncertain. A project can override it. | smart |
Project
Settings for the project of the open session. A value set here overrides the global one; x clears it again and the global setting applies. Without an open session the tab only says so.
| Setting | What it does |
|---|---|
| Verify command | What /verify and the verifier run
to check the work. Empty means verification is off for this project;
unset, Nemocode detects a command itself. |
| Home path | The folder the agent treats as its home. Unset, it is the project folder. |
| Context budget | Overrides the global context budget for this project. Steps through 32k, 64k, 128k, 200k and 400k, or back to the global value. |
| PUM loading | Overrides map / full for this project. |
| Ask before compaction | Overrides off / smart / always for this project. |
| Harvest external sessions | Facts from Claude Code and Codex sessions in this project become proposals. Off by default. |
| Context on external prompts | External agents get project memory context with every prompt. Off by default. |
Below them are lists you open with Enter:
- Allowed commands — shell commands that run without
asking (the same list as
/allow). - Denied paths — paths you answered “never” to in an approval, or added here yourself. The agent can neither read nor write them; lift one to be asked again.
- Sources — extra folders that belong to the project, with indexing on or off.
- Hooks, Scheduled jobs, Custom agents and Connectors — see Skills, MCP, agents.
- Crosstalk pairs — which sessions may message each other.
Display
Only what you see in the transcript. Nothing here changes what the agent does.
| Setting | What it shows | Default |
|---|---|---|
| Compact mode | One line per step, no detail blocks. | off |
| Recall details | What the mission recalled from the memory when it started. | on |
| Harvest details | Reindexing and what the curator harvested. | on |
| Intake events | How your request was understood, and clarifications. | on |
| Tool output | Result previews and edit diffs. | on |
| Model thoughts | A live preview of what the model is writing. | on |
| Preload / restore lines | The technical notes about preloading and a restored conversation. | off |
| Show $ cost | Money per mission, using the prices under Advanced. | on |
| Context % of budget | Count the context against your budget instead of the model's own window. | on |
| Show used model | Name the model a mission actually ran on. | off |
| Show cache reads | Cache-read tokens in the token line. | off |
| Context-trim notice | Say when old output was hidden to free context. | off |
Advanced
Tuning, and the switches that remove limits.
| Setting | What it does | Default |
|---|---|---|
| Context budget | How much context a mission may use before Nemocode compacts. Smaller is cheaper and forgets sooner. Steps through presets from 8k to 1M tokens, or Enter types your own value (below 4k counts as auto). The line shows the resulting number and where compaction starts — at 75 % of the budget. See Context & cost. | auto (the model's window minus a reserve for the reply) |
| Max iterations | Hard cap on how many tool rounds a mission may take before it gives up (1–5000). | 1500 |
| Goal max rounds | How many “not met yet” checks
/goal accepts before it stops working on the condition
(1–1000). Steps through 10, 20, 40, 100, 200, 1000, or
Enter types a number. | 40 |
| Ultra budget | A token cap for one /ultracode or
/ultrareview run. Steps through 500k, 1M, 2M, 5M, 10M and 20M
tokens. Nemocode warns at about 80 % and asks at the limit. | off (unlimited) |
| Read scope | Where the agent may read
(read_scope, NEMOCODE_READ_SCOPE): project keeps it
inside the project, ask asks for each outside folder, any lets it read
anywhere. Writing stays protected either way. Set it globally, per project
(Settings ▸ Project ▸ Read scope) or for one session (Read scope (this session)). | any |
| Read-only paths | Folders the agent may read but never change, also not through the shell . You add one in the path dialog (Read only — always) or add it yourself under Settings ▸ Project ▸ Read-only paths, where you can also remove it. | none |
| PUM links | Which projects may read each other's project memory: one list names the projects this one may read, the other the projects allowed to read this one. Both sides must agree, and answers from a linked memory are marked as hints from another project. | none |
| Trim threshold | When old tool output is hidden to save
context (trim_threshold_percent, NEMOCODE_TRIM_THRESHOLD): only when
the cache has gone cold or the conversation nears compaction, never while the cache is warm. Percent of the
compaction point, 0–95. | 80 |
| Diff context lines | How much of a diff the agent sees after editing (1–2000). | 100 |
| Pricing | The per-million prices the cost display uses for the current model. Enter picks from a live list or lets you set your own; x returns to the built-in or live price. See Models & cost. | built-in / live list |
| Prompt cache TTL | How long the provider keeps the prompt warm: 5m or 1h. Longer survives pauses but a cache write costs twice the normal input price instead of 1.25 times; switching it on asks for confirmation. Only Anthropic models use this setting; other providers manage their own caching. | 5m |
| CAST context trimming | How eagerly old tool output is dropped when it is cheaper to drop it than to carry it: off, careful, balanced, eager. Enter steps to the next one. Explained in Context & cost. | balanced |
| Compaction max cut | How much of the conversation one compaction may remove in a single step (10–100 %, in steps of 10). | 50 % |
| Compaction cooldown | How many rounds must pass after a compaction before the next one may start. | 3 rounds |
| GPC pre-compression | Shorten the carried conversation at mission start: on (always), off, or smart (decided per mission). See Context & cost. | off |
| Advisor | Let the agent ask a stronger model at decision points. It advises, it does not act. Three states: off; on; on with any model, which also allows an advisor that is not stronger than the main agent. The model is chosen under Models ▸ advisor. | off |
| Lazy tools | Offer tools on demand instead of all of them upfront. | off |
| Vision | Whether a pasted image reaches the model as an image: auto decides from the model, or force it on or off. | auto |
| Plugins | Off, only standalone skills are found — no skills that come bundled inside a plugin. | on |
| PUM (knowledge graph) | Off, missions run without the project memory: nothing is recalled, nothing is harvested. Meant for comparisons, not for daily work. | on |
| Permission rules | Opens the permission page (the same as
/permissions): mode, allow / ask / deny rules, block and
confirm lists. | — |
| Confine shell to project | Whether commands may touch anything outside your project. See Permissions & safety. | on |
| Dangerously skip all permissions | Nothing asks any more. Needs a confirmation to switch on. See Permissions & safety. | off |
| Dangerously allow all paths | Drops the project boundary for files and the shell. Needs a confirmation to switch on. | off |
Models
One line per role — the main agent, the verifier, the curator, the advisor and the rest. Give each its own model and provider, or leave it on the default. Enter picks the model, x returns the role to its default, e sets how hard that role thinks. All roles are listed on Models & cost.
The first line is Only this session. With it ticked, every model you pick is stored on this session alone — one session on security with a big model, the next on frontend with a fast one, and neither changes the other. Untick it and the session goes back to your global choice.
Usage & Stats
Usage (also what /usage opens) shows tokens in
and out, cache reads and writes, and what it cost, in five blocks:
Today across all projects, the Current session,
All time, By model and By role. The last one breaks
the cost down by role, so you see what the verifier, the curator and each
subagent (explore, delegate) cost next to the main agent. Below them, the last
five missions of this session with their tokens in and out. Money is computed
with the prices under Advanced ▸ Pricing; old missions show as unknown, and the
per-role rows are estimated from tokens.
Stats is the long view: tokens per day for the last 14 days, total tokens and API time, missions, sessions and projects.
Settings outside /config
A few settings live where you use them:
| Setting | Where |
|---|---|
| Provider and API key, local endpoint, custom endpoint | /connect |
| Main model, its context window | /model |
| Thinking effort | /effort |
| Permission mode | /mode, /permissions |
| Path limits of one session | /paths |
| Allowed commands | /allow |
| Which memory folders feed the PUM | /pum-folders |
| Skills, MCP servers, hooks, agents, schedule, connectors | /skills, /mcp, /hooks, /agents, /schedule, /connectors |
| Licence, update, export, health | /licence, /update, /export, /doctor |
Environment variables
Most settings can be preset from an environment variable. The rule is
simple: the variable only sets the starting value. As soon as you
change the setting in Nemocode, your choice wins. There are two exceptions where
the variable wins and the line in /config becomes read-only until
you unset it and restart: NEMOCODE_INDEPENDENT_VERIFIER (Verify
subagent) and NEMOCODE_PUM_SYNC_PUSH_INTERVAL_S (auto-push interval).
For the two external-agent switches, NEMOCODE_HARVEST and
NEMOCODE_PROMPT_CONTEXT in the process of a single
nemocode hook call win over the project setting for that call.
| Setting | Variable |
|---|---|
| Provider, model | NEMOCODE_PROVIDER, NEMOCODE_MODEL |
| API keys | ANTHROPIC_API_KEY, OPENAI_API_KEY,
OPENROUTER_API_KEY, MOONSHOT_API_KEY,
OPENCODE_API_KEY, OPENCODE_GO_API_KEY,
DEEPSEEK_API_KEY, GROQ_API_KEY,
CEREBRAS_API_KEY, OLLAMA_API_KEY |
| Ollama address | OLLAMA_URL |
| Custom endpoint | NEMOCODE_CUSTOM_BASE_URL,
NEMOCODE_CUSTOM_MODEL, NEMOCODE_CUSTOM_API_KEY,
NEMOCODE_CUSTOM_WINDOW |
| Effort | NEMOCODE_EFFORT |
| How long Nemocode waits for the browser sign-in with a ChatGPT account (default 300 seconds) | NEMOCODE_OAUTH_TIMEOUT_SECONDS |
| Prompt cache TTL | NEMOCODE_CACHE_TTL |
| Permission mode | NEMOCODE_PERMISSION_MODE |
| Skip all permissions / allow all paths / confine shell | NEMOCODE_DANGEROUSLY_SKIP_PERMISSIONS,
NEMOCODE_ALLOW_ALL_PATHS, NEMOCODE_CONFINE_SHELL |
| Verify subagent (locks) | NEMOCODE_INDEPENDENT_VERIFIER |
| Advisor | NEMOCODE_ADVISOR, NEMOCODE_ADVISOR_ALLOW_WEAK |
| Auto-approve plans, mode after approval | NEMOCODE_AUTO_APPROVE_PLANS,
NEMOCODE_PLAN_APPROVED_MODE |
| Ultra budget, goal max rounds | NEMOCODE_ULTRA_MAX_TOKENS,
NEMOCODE_GOAL_MAX_ROUNDS |
| CAST, GPC, lazy tools, vision | NEMOCODE_CAST,
NEMOCODE_CAST_MODE, NEMOCODE_GPC,
NEMOCODE_LAZY_TOOLS, NEMOCODE_VISION |
| Compaction max cut, cooldown | NEMOCODE_MAX_CUT_PERCENT,
NEMOCODE_COMPACT_COOLDOWN |
| Cross-session talk, plugins, remember tool | NEMOCODE_CROSS_SESSION_TALK,
NEMOCODE_PLUGINS_ENABLED, NEMOCODE_REMEMBER_TOOL |
Learn skills from missions (skill_review) | NEMOCODE_SKILL_REVIEW |
| Project memory, harvest mode, trigger | NEMOCODE_PUM_ENABLED,
NEMOCODE_PUM_HARVEST_MODE, NEMOCODE_HARVEST_TRIGGER |
| Embedding model, auto-pull, startup refresh | NEMOCODE_EMBED_MODEL,
NEMOCODE_EMBED_AUTO_PULL, NEMOCODE_EMBED_STARTUP_REFRESH |
| Push interval for team sync (locks) | NEMOCODE_PUM_SYNC_PUSH_INTERVAL_S |
| External-agent hooks | NEMOCODE_HARVEST, NEMOCODE_PROMPT_CONTEXT |
| Where new projects are created | NEMOCODE_PROJECTS_DIR
(default ~/NemocodeProjects; environment only) |
| Server port, server address for the terminal UI | NEMOCODE_PORT,
NEMOCODE_SERVER_URL |