Configuration
Every switch agent.toml accepts, with its default.
Model and web
| Key | Default | Meaning |
|---|---|---|
model |
picked for the key that is set | Any model id; anthropic/…, openai/…, gemini/… (or google/…), bedrock/… and you/… name the provider. See Quick start |
reasoning_effort |
the provider’s | How hard the model thinks. Claude takes low to max, OpenAI’s GPT models none to max, others low, medium or high |
max_tokens |
the provider’s | Cap on output tokens per model call |
models |
[] |
Other models a client may pick per session (the web chat’s picker, or model in the HTTP API) |
provider |
from the model id | you, openai, anthropic, gateway, gemini or bedrock; see Providers |
base_url, api_key_env |
the provider’s | Point the provider at another endpoint or key |
context_limit |
from the provider | Override the model’s context window |
vision |
from the provider | Whether the model can see images |
audio_model |
the session’s model if it hears, else the cheapest that does | The model that transcribes audio for read_file |
web |
picked for the keys that are set | client gives the agent web_search and web_fetch; server lets the model provider search; off |
search |
the first search key set, You.com first | Web search provider: you, parallel, firecrawl, context.dev, brave or searxng |
search_mode |
highlights |
You.com Search mode: highlights or snippets |
search_count |
5 | Results per search |
freshness, include_domains |
Passed to You.com Search | |
exclude_domains |
[] |
Never search or fetch these domains |
exclude_terms |
[] |
Withhold results and pages containing any of these strings |
exclude_echo |
false |
Withhold results and pages that quote the task word for word |
exclude_terms and exclude_echo exist for evals whose answers are republished online. With exclude_terms, list the task ids or the dataset name and the agent cannot read an answer key by accident. exclude_echo covers copies no list can name: a page that repeats twelve or more words of the task in a row is withheld, because a page that reproduces a benchmark question usually prints its answer. Words the task puts in quotation marks, such as a title, don’t count.
Tools and turns
| Key | Default | Meaning |
|---|---|---|
tool_style |
typed |
bash leaves only the bash tool |
batching |
true |
Several tool calls per turn; safe ones run concurrently |
planning |
false |
Adds the plan tool |
memory_max |
0 (off) | Adds the memory tool, capped at this many bytes |
index |
false |
Adds search_files: full-text search over the workspace |
approval |
ask |
ask waits for an answer before any tool not marked safe (bash, file writes, custom and MCP tools); once asks the first time each tool is called in a session and remembers a yes; never runs them without asking |
ask |
false |
Adds ask_user: the agent puts a question to the user and waits for the answer |
mid_run |
steer |
A message sent during a run: steer hands it to the model at its next step; queue holds it until the current answer is done |
max_turns |
60 | Per-session turn limit |
max_cost |
0 (off) | Per-session limit in dollars |
bash_timeout |
600 | Seconds a bash command may run in all |
bash_wait |
60 | Seconds a bash call waits for its command before returning |
bash_background |
true |
A command still running after bash_wait keeps going as a background job; false stops it |
read_before_write |
true |
write_file refuses to overwrite a file the agent hasn’t read, or that changed since |
mcp_tools |
auto |
listed shows every MCP tool to the model; deferred puts them behind tool_search and tool_call; auto defers above 20 tools |
profile |
true |
false ignores profiles/<model>.* |
Context
| Key | Default | Meaning |
|---|---|---|
compaction |
true |
Truncate long tool output as it enters the conversation |
tool_output_max |
20000 | Characters of command output kept, from its start and end; the full text is saved in .egin/out/ in the sandbox |
page_max |
30000 | Characters of a page, document or search result shown at once. Longer pages and files come a part at a time |
elision |
true |
Stub out old tool output once the context fills |
elide_at, keep_recent |
0.5, 6 | Start at this share of the limit; keep this many recent results whole |
summarisation |
true |
Summarise old turns |
summarise_at |
0.75 | Start at this share of the limit |
Sandbox
| Key | Default | Meaning |
|---|---|---|
sandbox |
auto |
auto, microsandbox, e2b, vercel, daytona, daytona-macos, docker, docker-cloud or local |
image |
python:3.13-slim |
The VM image |
cpus, memory_mb |
2, 2048 | VM size |
idle_timeout |
600 | Seconds before an idle VM stops |
workdir |
Working directory inside the VM | |
network |
["*"] |
Egress: ["*"] all, [] none, or a list of domains |
secrets |
[] |
Host secrets substituted only on requests to a named host |
See Sandboxing for how each backend behaves.
Per-model profiles
Models differ. One follows typed tools well, another does better with only bash. Put switches that suit one model in profiles/<model>.toml; they apply whenever that model runs, under agent.toml.
# profiles/muse-ember-1.0.toml
tool_style = "bash"
batching = false
egin evolve writes these files for you from eval results. See Evals.