Reference

Configuration reference

Every setting that Glider reads, with its type, its default and its function. Each key here is a field in internal/config/config.go. A key that this page does not list does not exist.

Which file. Glider reads configs/glider.yaml from the current directory. To use a different file, give --config: .\glider.exe --config configs\glider.local.yaml. Refer to the profiles below.

The profiles that Glider supplies

File	Function
`configs/glider.yaml`	The default. A hybrid: local models and cloud, with transparent interception on.
`configs/glider.local.yaml`	Pure local. No cloud key. Refer to Pure local.
`configs/glider.cloud.yaml`	Cloud only.
`configs/glider.transparent-test.yaml`	A profile for a test of transparent interception.

What applies immediately, and what needs a restart

Glider applies most changes while it operates. Edit the file, or save from the Config page of the dashboard.

Applies immediately	Needs a restart
The routing rules, the model aliases, the context thresholds, the log level, and the clients for the backends and the models	The listen ports, the CA of the MITM, the host list, and each setting for transparent interception

To see the condition: the Config page of the dashboard shows the list of hot-swap modules. A response also carries the header X-Glider-Backend-Reload after you save from that page.

server

Key	Type	Default	Function
`proxy_port`	int	`8080`	The gateway. It gives the `/v1/*` endpoints.
`dashboard_port`	int	`8081`	The dashboard, the `/api/*` endpoints and this documentation.
`log_level`	string	`info`	`debug`, `info`, `warn` or `error`.

thresholds

Key	Type	Default	Function
`max_local_context_tokens`	int	`8000`	The ceiling on the context that Glider gives to a local model.
`idle_unload_timeout`	duration	`5m`	Glider removes a model from memory after this time with no use.
`request_timeout`	duration	`10m`	The HTTP timeout for a local backend. Increase it for a large model with a long tool loop.
`default_max_tokens`	int	`8192`	The `max_tokens` of a completion. Ollama receives this value as `num_predict`. Increase it if a report stops in the middle of a sentence.

vram

Key	Type	Default	Function
`strategy`	string	`dynamic`	`static`, `dynamic` or `hybrid`. How Glider gives the GPU memory to the models. The supplied `glider.yaml` sets `hybrid`.
`headroom_mb`	int	`512`	The memory that Glider keeps free, in MB.
`max_loaded_models`	int	`3`	The count of models that stay in memory together.
`gpu_assignments`	map	—	A model name to a GPU index. `PUT /api/gpu-assignments` writes this map.

models

A list. Each entry describes one model.

Key	Type	Default	Function
`name`	string	—	The name of the model, as its backend knows it.
`backend`	string	`ollama`	The name of an entry in `backends`.
`vram_estimate_mb`	int	`5000`	The memory that this model needs. The strategy for the VRAM uses this value.
`max_context`	int	`8192`	The context window of the model.
`capabilities`	list	—	Labels such as `code`. A rule can select a model by a label.
`adapter`	string	—	The name of a LoRA adapter to apply. vLLM only.
`adapters[].name`	string	—	The name of one LoRA adapter that Glider can load.
`adapters[].path`	string	—	The path to that adapter.
`keep_warm`	bool	`false`	Keeps the model in memory. The idle timeout does not apply to it.

model_aliases

A map from a name that a CLI sends to a model that you have. This is how a request for gpt-4o reaches a local model with no change in the CLI.

model_aliases:
  claude-3.5-sonnet: qwen2.5-coder:14b
  gpt-4o: qwen2.5-coder:14b

routing

Refer to Routing for the full priority sequence and each trigger type.

Key	Type	Default	Function
`rules`	list	—	The rules. Each field is here.
`default`	string	—	`local` or `cloud`. It changes the target of the priority-0 rule. Therefore a profile becomes pure local with no edit to each rule.
`default_local_model`	string	—	The model for that rule when `default` is `local`.
`allow_cloud_fallback`	bool	`true`	With `false`, Glider never goes to the cloud after a local model fails. Pure local uses `false`.
`complexity_from`	string	`heuristic`	`heuristic`, `cursor` or `both`. It selects the source of the complexity score.
`turn_family_ttl`	duration	`90s`	How long a turn family stays open for a summary or a title that comes after it. Refer to MITM.

routing.task_classifier

It adds rules that use a heuristic. The rules go between the explicit commands and the context size.

Key	Type	Default	Function
`enabled`	bool	`true`	Turns the classifier on.
`tools_force_cloud`	bool	`true`	A request that has `tools[]` goes to the cloud. Set `false` to permit a local model with tools. That model must support tools.
`must_cloud_patterns`	list	built-in	Regular expressions that force the cloud. A value here replaces the built-in list.
`small_local_patterns`	list	built-in	Regular expressions that select the local model. A value here replaces the built-in list.
`local_model`	string	—	The model for a match with a small task.
`cloud_backend`	string	—	The backend for a match that must use the cloud.
`cloud_model`	string	—	The model for that match.
`tools_priority`	int	`85`	The priority of the rule for the tools.
`must_cloud_priority`	int	`80`	The priority of the rule that must use the cloud.
`small_local_priority`	int	`70`	The priority of the rule for a small task.

routing.complexity

Key	Type	Default	Function
`enabled`	bool	`false`	Adds a rule that uses the complexity score.
`cloud_above`	int	`70`	A score of this value or more goes to the cloud. The range is 0 to 100.
`priority`	int	`75`	The priority of that rule. It is below must-cloud (80) and above small-local (70).
`cloud_backend`	string	—	The backend for the rule.
`cloud_model`	string	—	The model for the rule.

routing.tool_followup

It controls where the calls of a tool go, after a parent turn has a route.

Key	Type	Default	Function
`enabled`	bool	`true`	Turns this control on.
`inherit_parent_default`	bool	`true`	A call starts with the decision of its parent turn.
`reevaluate`	bool	`true`	Permits a safe tool to go to the local model, also when the parent went to the cloud.
`local_tool_allowlist`	list	—	The names of the tools that can go to the local model. An example is `read_file` or `grep`.
`cloud_tool_denylist`	list	—	A match here forces the cloud. An example is `Shell` or `Write`.

orchestration

Key	Type	Default	Function
`fan_out.enabled`	bool	`false`	Permits one message to go to more than one worker. The supplied `glider.yaml` keeps this section as a comment.
`fan_out.max_workers`	int	`2`	The count of workers for one fan-out.
`concurrency.worker_queue_size`	int	`16`	The size of the queue of the workers.
`concurrency.result_chan_size`	int	`32`	The size of the channel for the results.
`concurrency.max_inflight`	int	`4`	The count of requests that operate together.

orchestration.tools

Refer to Tools & MCP for each tool.

Key	Type	Default	Function
`workspace`	string	`~/.glider/workspace`	The root of the sandbox for `fs_`, `git_` and `shell_exec`. Do not set it to `.`, or the tools operate in the repository of Glider.
`allow_shell`	bool	`false`	Permits `shell_exec`. It is off, because the tool is dangerous.
`shell_allowlist`	list	—	The commands that `shell_exec` can run.
`allow_hosts`	list	any	Limits `http_fetch` and `web_fetch` to these hosts. An empty list permits each host.
`web_search.provider`	string	`auto`	`auto`, `duckduckgo`, `brave`, `tavily`, `serpapi` or `searxng`.
`web_search.max_results`	int	`5`	The count of results. A value of more than `10` becomes `10`.
`web_search.brave_api_key_env`	string	`BRAVE_SEARCH_API_KEY`	The environment variable that holds the key for Brave. `BRAVE_API_KEY` also operates.
`web_search.tavily_api_key_env`	string	`TAVILY_API_KEY`	The environment variable that holds the key for Tavily.
`web_search.serpapi_key_env`	string	`SERPAPI_KEY`	The environment variable that holds the key for SerpAPI.
`web_search.searxng_url`	string	—	The URL of your own SearXNG. The variable `SEARXNG_URL` also operates.
`web_search.fetch_max_bytes`	int	`65536`	The ceiling on the body that `web_fetch` reads. A value of more than `262144` becomes `262144`.

cloud

Key	Type	Default	Function
`providers[].name`	string	—	The name that a rule uses, such as `openai` or `anthropic`.
`providers[].api_key_env`	string	—	The environment variable that holds the key. Glider never writes the key in the config.
`providers[].base_url`	string	—	The endpoint of the provider.
`rate_limit.requests_per_minute`	int	`30`	The ceiling on the requests to the cloud.
`rate_limit.tokens_per_minute`	int	`100000`	The ceiling on the tokens to the cloud.
`budget_cap_usd`	float	`50`	The ceiling on the cost.

backends

Key	Type	Default	Function
`name`	string	—	The name that `models[].backend` uses.
`type`	string	—	`ollama`, `vllm`, `openai` or `anthropic`.
`url`	string	—	The address of the backend. Ollama uses `http://127.0.0.1:11434`.
`health_check_interval`	duration	`30s`	How often Glider tests that the backend operates.

dashboard

Key	Type	Default	Function
`enabled`	bool	`true`	Turns the dashboard and the `/api/*` endpoints on.

transform

It controls the messages that Glider gives to a local model.

Key	Type	Default	Function
`enabled`	bool	`false`	Turns the transformation on.
`trim_context`	bool	`false`	Removes the older messages before Glider sends the request.
`augment_prepend`	string	—	Text that goes before the prompt.
`augment_append`	string	—	Text that goes after the prompt.
`local_context`	string	`latest_turn`	`latest_turn` sends the most recent turn only. `full` sends the messages with no limit.
`local_system_max_chars`	int	`4000`	The ceiling on the system content under `latest_turn`.
`local_episode_count`	int	`3`	The count of earlier episodes that go in the pack. A value of 0 stops this.
`local_episode_max_chars`	int	`1500`	The ceiling on the full block of the episodes.

mitm

Read MITM & interception before you change these settings.

Key	Type	Default	Function
`enabled`	bool	`true`	Turns the proxy of the MITM on.
`port`	int	`8082`	The listener for CONNECT.
`ca_cert`	string	`~/.glider/mitm/ca.crt`	The certificate of the CA of Glider.
`ca_key`	string	`~/.glider/mitm/ca.key`	The private key of that CA.
`hosts`	list	see the file	The hosts that Glider decrypts. Glider sends each other host with no decryption.
`passthrough_default`	bool	`true`	A host that is not in the list goes to its origin with no change.
`transparent`	bool	`true`	Turns interception at the level of the operating system on. It needs Administrator on Windows, or root on Linux.
`transparent_port`	int	`8083`	The ingress for the redirected connections. It is not the same as `port`.
`transparent_ports`	list	`[443]`	The destination ports that Glider intercepts.
`windivert_dll_path`	string	`~/.glider/mitm/windivert/WinDivert.dll`	Windows only. `WinDivertNN.sys` must be in the same directory. With no value, `transparent` does nothing.
`origin_on_local_error`	bool	`true`	If the local model fails, the request goes to its origin. Set `false` for pure local, so an error is visible.
`require_local_healthy`	bool	`false`	Glider selects the local model only when a health check succeeds. This stops a silent change to the origin.
`debug_agent_rpc`	bool	`false`	Adds logs, the dumps of the bodies, and `GET /api/mitm/debug/recent`. The variable `GLIDER_MITM_DEBUG_RPC` also turns it on.
`debug_dump_dir`	string	`~/.glider/mitm-debug/`	Where those dumps go.
`agent_rpc_fulfill`	bool	`true`	Permits a local reply on the Agent RPC plane of Cursor.
`agent_rpc_canned_on_error`	bool	`false`	Writes fixed text when the local model fails, in place of the origin.
`agent_rpc_canned_text`	string	`pong from glider (canned Path B)`	That text.
`agent_rpc_tool_codec`	bool	`false`	Permits a local reply for the loop of a tool. The variable `GLIDER_MITM_AGENT_RPC_TOOL_CODEC` also turns it on.

context

Refer to Context for the two layers and the selection.

Key	Type	Default	Function
`max_events`	int	`4096`	The size of the event ring in memory.
`retain_days`	int	`14`	How long the files `events-*.jsonl` stay on disk.
`warm_load_days`	int	`2`	How many days Glider reads again when it starts.
`background.token_budget`	int	`20000`	The ceiling on the block that goes with each delegate call.
`background.max_entries`	int	`20`	The count of entries in that block.
`background.max_entry_chars`	int	`4000`	One entry, before Glider removes its middle.
`summary.enabled`	bool	`true`	Permits a model to write the summary. With `false`, Glider still compacts, with the deterministic digest.
`summary.chain`	list	`[origin, cloud, local]`	The order that Glider tries. The values are `origin`, `cloud`, `local` and `none`.
`summary.origin_vendor`	string	—	Which CLI answers for `origin`. With no value, Glider uses the front CLI of that workspace. Only a headless template can do this work.
`summary.cloud_backend`	string	—	The backend for `cloud`.
`summary.cloud_model`	string	—	The model for `cloud`.
`summary.local_model`	string	`routing.default_local_model`	The model for `local`.

The environment variables

Glider reads .env and .env.local from the repository root when it starts. It does not change a variable that your shell already gives. Both files are in .gitignore. Copy .env.example to begin.

The switches for MITM

Each one accepts 1, true, yes or on, and the case does not matter. A switch only turns a setting ON. It never turns one off, therefore you cannot use it to remove a value from the config file.

Variable	The setting that it turns on
`GLIDER_MITM_DEBUG_RPC`	`mitm.debug_agent_rpc`
`GLIDER_MITM_AGENT_RPC_FULFILL`	`mitm.agent_rpc_fulfill`
`GLIDER_MITM_AGENT_RPC_CANNED`	`mitm.agent_rpc_canned_on_error`
`GLIDER_MITM_AGENT_RPC_TOOL_CODEC`	`mitm.agent_rpc_tool_codec`

GitHub, for MCP

Variable	Function
`GITHUB_PERSONAL_ACCESS_TOKEN`	A personal access token. `GITHUB_TOKEN` and `GH_TOKEN` also operate.
`GLIDER_GITHUB_OAUTH_CLIENT_ID`	The id of your OAuth app. It turns the Sign in with GitHub button on.
`GLIDER_GITHUB_OAUTH_CLIENT_SECRET`	The secret of that app. With no secret, Glider uses the device procedure.
`GLIDER_GITHUB_OAUTH_SCOPES`	The scopes to ask for, such as `repo read:org read:user`.
`GLIDER_GITHUB_OAUTH_REDIRECT_URI`	Changes the callback. The default is `http://127.0.0.1:8081/oauth/callback`, and the OAuth app must give exactly this value.

Web search

Variable	Function
`BRAVE_SEARCH_API_KEY`	The key for Brave. `BRAVE_API_KEY` also operates.
`TAVILY_API_KEY` · `SERPAPI_KEY`	The keys for those services.
`SEARXNG_URL`	The address of your own SearXNG.

A key of orchestration.tools.web_search names the variable to read. Therefore you can use a different name. Refer to orchestration.tools.

Other

Variable	Function
The variable in `cloud.providers[].api_key_env`	The key of that provider. An example is `OPENAI_API_KEY` or `ANTHROPIC_API_KEY`.
`GLIDER_HOME`	Moves the directory of the credentials from `~/.glider/credentials` to `<GLIDER_HOME>/credentials`. It changes only the credentials.

What Glider tests in a config

Glider tests the config when it reads it. POST /api/validate does the same tests, and it applies nothing. A test gives an error or a warning. An error stops the config. A warning does not, because the condition can be correct: a model can exist in a backend that Glider did not scan.

A value must be in the set

Key	The values that it accepts
`server.log_level`	`debug` · `info` · `warn` · `error`
`vram.strategy`	`dynamic` · `hybrid` · `static`
`routing.default`	`local` · `cloud`
`routing.complexity_from`	`heuristic` · `cursor` · `both`
`transform.local_context`	`latest_turn` · `full`
`routing.rules[].trigger.type`	`explicit` · `regex` · `context_size` · `always` · `script` · `composer_wrapup`

routing.rules[].action.target is necessary. A value that is not local or cloud gives a warning, and not an error.

A value must be in a range

Key	Range
`server.proxy_port`	1 to 65535
`server.dashboard_port` · `mitm.port`	0 to 65535
`routing.complexity.cloud_above`	0 to 100
`thresholds.max_local_context_tokens` · `vram.headroom_mb` · `vram.max_loaded_models` · `models[].vram_estimate_mb` · `models[].max_context`	0 or more
`thresholds.idle_unload_timeout` · `thresholds.request_timeout` · `routing.turn_family_ttl` · `backends[].health_check_interval`	A duration that Go can read, such as `90s`, `5m` or `10m`.

A name must exist, and it must be one name

Rule
`backends[].name` and `backends[].url` are necessary.
`models[].name` is necessary, and two models cannot share a name.
An empty key of `model_aliases`, or an empty target, is an error.
The GPU index in `vram.gpu_assignments` must be 0 or more, and it must be less than the count of the GPUs that Glider found.

These give a warning only

Condition
`models[].backend` does not name an entry in `backends`.
A target of `model_aliases` is not in `models`.
A model in `vram.gpu_assignments` is not in `models`.
A rule sends work to a local model that is not in `models`.
A model is not in the list that Glider found in the backend. This is an error when the test is not soft.

A trigger must carry its own field

Type	The field that it must have
`explicit`	`commands`
`regex`	`pattern`, and the expression must compile
`script`	`file`
`context_size`	`operator`

routing.rules[].name is necessary for each rule, and trigger.type is necessary. The patterns of routing.task_classifier must compile as regular expressions.

How to test a config

GET  /api/config              # the config that operates now
GET  /api/config?format=yaml  # the same, as YAML
POST /api/validate            # tests a config, and applies nothing
GET  /api/router/lint         # finds an error in the routing rules

The Playground page of the dashboard reads a message with the same parser that Glider uses, and it changes nothing.

Next: Routing gives each trigger type, Context gives the two memory layers, and the API gives each endpoint.